> A chatbot account tied to Alfozanās spouse, which shared identifiers with his own account, submitted prompts between Sept. 24 and Sept. 30 about building, installing, obscuring, and remotely accessing a recording device, prosecutors said. Queries included how to ābuild a continuous audio recorderā and āencrypt the SD card,ā prosecutors claimed. On Sept. 27 and 28, the account asked about the specific device that was later found, including its unique MAC address.
We had a very public case in south of Sweden where a woman asked chat-gpt last year about how a failed rape attempt would look like and where in the local park there would be spots without any cameras, a full day before she reported to have been followed and attacked in that park. Quite recently the court found her guilt of filling a false police report, citing among other things the chat log (the police spent over 400 work hours on the initial report).
It is the first case I seen in the news of LLM logs being used as evidence.
There was a mom in Utah who killed her husband poisoning him with a moscow mule spiked with fentanyl who googled "luxury prisons for the rich" as well as "how long does life insrance companies takento.pay" "can you delete everythjng on an old icloud account" "can deleted text messages be retrieved" "what information can be cbtained from a cellphone" "how to turn find my iphone off" "can cops force you to do a lie detector test," "when does FBI get involved in a case," "what is considered a lethal.dose of fentanyl" and "what are you allowed inside utah jails."
She took out $2 million in life insurance on him and was having an affair with their handyman and they were having disputes about $$.
She tried to poison him twice and he stayed with her. Now, thats ... uh ... love.. i think. She must have been one top notch narcissist to drag him along.
An essay it seems like, which wasn't included in the trial even:
> Crampton Brophy's how-to treatise detailed various options for committing an untraceable killing and professed a desire to avoid getting caught. Circuit Judge Christopher Ramras ultimately excluded the essay from the trial, noting it was published in 2011. A prosecutor, however, alluded to the essay's themes without naming it after Crampton Brophy took the stand.
All other things being equal, why would it matter if a suspect instead of saying something obviously incriminating, had instead said something diametrically opposed? Is that what you are asking?
If so, I'm confused. Why wouldn't it matter? Wouldn't this necessarily only be the case where the Google search doesn't really matter at all (i.e. there is sufficient other incriminating evidence).
Anyway I also think you're missing what was merely meant to be a comment in jest. But now I'm actually confused what point you are trying to make.
If the implication is the Google search is an implicit confession, because in a Google search you think you are not being watched, and thus are confessing your honest desires, why would his search of love be a "claim of innocence" rather than a "confession of love"?
> Similar to how people realized their Google searches were accessible to law enforcement.
There's probably a regression in beliefs, actually. Even here on HN, for more tech-savvy people.
We used to know all the hard lessons of lack of privacy in browsing the internet, using Google services, etc. I remember when (tech) people were up in arms that Gmail scanned your emails, remember that?
Well, all of that is seemingly forgotten in the age of LLM chat bots. It's back to square one. People need to relearn all the lessons, it's as if AI magically erased all of what we used to know about online privacy.
I started asking things I'd normally want no one to know about (but if they do, it's fine) in private browser windows, not logged in (duck.ai most of the time anyway), usually VPN toggled ON just for that. Now I just use an offline model for that but since RAM is the bottleneck on my mac so to even run a 6-7GB model I have to start aggressively killing anything and everthing that's running and then hope to clear up enough on this 16GB RAM machine. So I am thinking of hosting one on one of the oracle free tier instances (if I can somehow use that whole 24GB RAM, that used to be, at once ā not sure that's available anymore). I tihnk Kaggle also offers something like that.
Also makes it important to not let other people use your LLM account... Could have more far-reaching consequences than having to make good on the offer to buy doughnuts for the office^
^being the classic message that gets sent from the account of someone who didn't lock their computer when they walked away from their desk.
This means no business should use AI providers (Cloud AI). Businesses do a LOT of shady stuff and all of it will be available. And, their IP or special sauce is all available to Anthropic/OpenAI to sell to other companies as AI solutions. Just like how AI solved recent math problems.
Local LLMs have the potential to convert computer use into what most people already thought it was: Private. In this, they could be one small step toward pushing back the Borg.
Note to EFF and Cato: You each do yeoman's work for privacy in the realm of policy and law. Keep it up, and add this--relentless encouragement of local AI. Rights and progress are as often driven by technological realities as by enlightened policy and law.
This is why a lot of companies that use Ai negotiate ZDR terms with their model providers of choiceā¦ Using local LLMs is a choice, but not one a lot of companies can support nimbly at scale.
Before AI tools like Plaud.AI, this wasnt near possible because of work/time, and no person usually do not remember exactly who said what 6 month before,unless there is some email conversation.
With Plaud, you can check unlimited backwards who said what in which meeting and what was wanted to achieve - very interesting actually to find things in there 6m later, while people said they will do it at ends of month and youself just forgot it already - and these systems allow you to chase exactly these discrepancies.
Before, they'd hire some humans to do the research. No self respecting C-suite person would ever consider doing something resembling actual work. How long have corps been required to keep copies of emails/chats/etc?
This has been happening long before LLMs, and this eBay case is notable only for the fact that it was actually publicly documented and prosecuted, not for the fact that it happened. I've personally been privy to two situations with other companies where the owner was doing something like this but nothing came of it legally. See also Meta employees looking up nudes streamed from people's Meta spy glasses, and Flock-using police stalking their exes. Smaller scale, but this also happens all the time with internet community moderators/admins. Humans are on average cretins, give them authority and the ability to abuse it and they will abuse it.
A good chunk of what keeps openai and similar companies afloat is subpar people who are willing if not eager to do crappy stuff to other people. A lot of the world works this way actually.
The only word for it is hubris. The overwhelming sense that your ability in one area has some magical transitive property to ability in another.
Can I write code? Sure. Im a bad designer, I can follow along in accounting, Im a terrible actor.
I would guess that there is much more to this story than just someone "bugging" a conference room. Though I would venture to guess that the keystone cops never get to the bottom of it.
Or, more accurately, people make stupid mistakes when they lack the experience. Intelligent people often succumb to the fallacy that because they can learn things quickly, they can skip the learning part just figure things out as they go.
Even my non tech elderly parents know that. The fact that they used LLMs to figure out how to best bug the executives really shows how poor their judgement is.
The weird thing is that I've had other HN users argue with me and claim that LLM chat users should be entitled to privacy, as if it were a priest or lawyer.
Paul staying on X was a real failure of judgment imo. He made the right call and then backed out of it. I don't remember if they invited Elon to startup school before or after, either way I don't think that that is the kind of connection you want.
Yup and thatās just ONE guy. The fact that I was able to bring an example on HN of literally the guy who founded YC, imagine what chance the rest of the āplebsā have to resist the siren song.
If you want to know what I think ā¦ they make both parents work long hours for corporations, stick their kids in government schools, which are run like prisons[1] where now if they act too rambunctious they are diagnosed with ADHD and prescribed amphetamines. Meanwhile the elderly parents who took care of you when you were little are put into nursing homes because their adult children have no time to take care of them either. The amount of medications for the elderly in nursing homes is even greater than for their children. The teenage girls have record levels of depression, coinciding with the rise of Instagram and TikTok. But they call it āsadnessā because they donāt want to link it with clinical depression, which is also quite high. Men are on opiates while one in five middle aged women is on antidepressants. And many people are now glued to their latest feed of doomscroll or video AI slop. Which is even worse than beforeā¦
So yeah. If you think the software is ābad for our kidsā, you should look at what it does to the adults. Endless arguing online with strangers for no reason. Maybe a bot will read it. Maybe a random person will reply or downvote it. Either way it wins because the system needs profits and the algorithm will prioritize āengagementā, which means one-sided outrage narratives that triggers someone to write an angry rant. We donāt really need those user-generated arguments anymore because AI can generate those, but then again we wonāt need humans to do most things online at allā¦
(Ironically, the above was an angry rant too, to open some strangersā eyes on HN, a social forum. But once you see it, you canāt really unsee it.)
He was inching into anti-dei culture war stuff so I'm not sure how much he actually disagrees with Musk's political goals so much as he rightly observers that Musk and Trump are complete maniacs.
Just to be clear, the difficult choice is not whether to use OpenAi/Anthropic/Grok/... versus a local LLM.
The choice is whether to hang out with people with whom you can't be sure they don't have a Plaud/Pocket/.. in their literal pocket which is continuously listening.
Unfortunately the set of such people will eventually be shrinking more and more until eventually only a few tinfoil folks are left. Dystopian? Just think back w.r.t. continuous tracking by Google/Facebook/.. Unthinkable 30 years ago, total standard today. Try to count the fraction of your emails that don't actually reside on a MSFT or Google server because the other end has an account there. That fight is over. You can just as well join them.
But in all seriousness ā no need for nudity, electronics canāt really get wet so it could be an indoor pool party where you finally let loose and discuss all your opinions about Putin.
Prepare to have that dug up 20 years in the future during some court proceedings.
Or at the border where it is already routine to ask for social media accounts (especially foreign nationals) you might someday five the border agent your login.
Seems like they select for people as dependent on their product as their customers. Maybe it helps them more to have true-believers rather than intelligent people.
> Spies used to put effort into these sorts of things
Assuming this is espionageāfor which I have seen zero evidenceāisnāt Saudi Arabia notorious for having a fuckup of an intelligence operation? The literal hatchet job that was the Khashoggi murder comes to mind.
> why bother? Nobody will do anything and the US president can be bribed
To exfiltrate OpenAI's executive-level planning? Probably not. I doubt he has that access anymore, what with the IC's decimation over the last two years.
In the 80s one of my high school teachers said her husband was an engineer who worked on surveillance technology and described how windows could be used as recording devices. People's voices inside caused the window to have minute vibrations, which could be measured by a laser pointing at the window's exterior and reflected back and then turned into audio output. Essentially, turning a window into a microphone diaphragm. It blew my mind.
I just looked it up, the technique was apparently invented by a Soviet engineer 40 years prior - Leon Theremin, who is also known for the musical wand heard on Led Zeppelin II and various 1960s horror movie soundtracks.
Vaguely similar, since electrical wires emit EM fields that are disturbed by human bodies, you can use a special camera to see who's in a building based on those disturbances. I'm not sure how high-fidelity or easy to spoil it is, but this was real technology at least 15 years ago.
This is a feature in modern wifi routers and every single one is capable of detecting presence of multiple people within an apartment. With some training, movement can be distinguished from the data.
Surely your conclusion should be that this was not the work of a spy - not a state sponsored one anyway. It's much more logical to conclude that he was working as a whistleblower or under some other personal agenda.
That's what you get when you make a dumb prompt and accept the first result LLM gives you. You have to push it to think out of the box, get into the realm of insane but plausible projects :).
That's because in all likelihood he is not a spy. He is an agent (lol, not an AI agent) who has a handler, who is the trained specialist[1]. The agent is disposable. The handler or case officer usually but not always has diplomatic cover courtesy of a nation state.
I couldnāt say whether this was state sponsored espionage or something else. I donāt see any indicators pointing one way or the other.
I see a lot of people mentioning financial motivation, and that decisions made in the room āmoves marketsā. Ugh, perhaps trueā¦and I have something to tell ya guys - economic espionage is a thing.
It sounds like he didnāt have great placement and access. I mean sort of, but not really. Janitors are much better.
Assuming he was, he could of course gone rogue to impress his handler. If directed by his handler, they arenāt very skilled. Any of that could be true.
Iām reminded of the idiot DIA officer that mishandled his own phone while being questioned and dumped incriminating evidence to them right on camera in the interrogation room. Geniuses.
The effect of getting caught is only half the story there; IQ is more generally correlated with all manner of "low time preference" type decision making.
Not really. Almost definitionally (because of the way we've built society over time, not by semantics or chance) crimes are a "shortcut" compared to the honest alternative.
throwing a prompt at an agent is often faster than cracking out man, lsusb, lsblk and friends, doesn't mean he can't do it without.
if I send a prompt to make an assortment of alpha numeric directories or whatever other braindead-to-automate-with-wildcards-and-ranges thing it doesn't mean I don't know how to do it, it means i'd rather think in English than regex and bourne shell and the extra few seconds and the token spend is worth the luxury of avoiding them for me, at that time, for that task.
I'll counter punching down on frontend engineers by positing that he was responsible for sandboxing their training infrastructure. Since clearly they don't know how to lock down an environment either without using llms
Someone who's objectively closest to this business should know the tracability of this. I'm shocked he didn't run a local LLM for this kind of sensitive work.
He doesn't seem like an idiot either, his background shows he's been doing work in the field since 2012.
I do wonder how much he was paid to throw away what was without a doubt a 1% career.
The fact that he used ChatGPT to plan and implement this wasn't how he got caught, though. He got caught because he put a physical device in a conference room, and they found it and subsequently put a camera in the room, etc. His career was already over at that point. I think wonking out over how he could have better opsec or whatever is misplaced - it wasn't that a red light blinked at OpenAI, "employee's wife is having suspicious chats!" - they just thought, after having caught him red handed, to have a look at his ChatGPT account and linked accounts.
Itās the criminals that you donāt catch that you have to worry about.
The fact that even this incompetent dude most likely still made off with some secrets means Iām sure some better trained ones got/are continuing to get a few more.
I donāt think this guy was operating under a spy network, he was probably just trying to make bank with insider info. He is screwed royally for his dumb move and nobody will come to his rescue. He got greedy.
They could have just taped it to the inside bottom of a full tissue box, and as long as they got it before the tissues were mostly empty I bet it would go undetected. Put a hand sanitizer next to it for extra camo.
Or hijack the TV's already always on mic. Just forward a copy of the audio to your server in addition to the one sent to Samsung or LG. WebOs and Tizen shouldn't be that hard to crack when you have access to unrestricted frontier models.
"OpenAI personnel found the device Sept. 30. It was plugged into a USB port on the back of a television in an executive conference room and had been concealed in toilet paper,"
I wonder what the IQ of someone needs to be to think that covering it up with tp makes it less suspicious instead of more. And I also wonder what the hiring process was like to hire someone with that mental frame.
should have enclosed it in a shell to look like a crappy USB A to thunderbolt adapter that some exec bought at an airport kiosk and left behind in the tv
The bar @ OAI is surprisingly low (watch a few minutes of their video presentations) - I've concluded it has something to do w how their recruiters define 'AI Native' (who assign their own meaning to meaninglessness of it), resulting in higher aggregate brain atrophy (hard to ignore once you notice it)
> And I also wonder what the hiring process was like to hire someone with that mental frame.
Remember, this is only the company responsible for potentially building and aligning AGI. No need to be setting high expectations toward their sandboxing or hiring capabilities.
Especially since you an just buy one, like Plaud and tape it to the underside of the desk, surreptitiously, instead of suspiciously accessing the back of the TV.
If there's one thing I've learnt about criminals (and spies), it's that a large percentage of them are incredibly careless and arrogant, often bordering on idiotic.
Remember, we've seen criminals caught because of such amazing decisions as:
- Writing and publishing books and essays describing their own crimes
- Adding hints into livestreams like they're trying to be the Riddler from Batman
- Searching for info on how to commit crimes via Google and not even bothering to clear their browser history
- Outright recording their crimes for YouTube/TikTok or livestreaming them on a site like Twitch or Kick.
The average criminal seems to be careless to the point of insanity.
That said, I guess it's also probably because the average person doesn't seem to realise they're being recorded if they use LLMs or search engines, and the average criminal or spy is no more tech savvy than that.
I am not sure whether it bodes well for humanity that the advice "If you want to break the law use local abliterated model or at least a Chinese one" actually has to be written down.
Recall that multiple of the Chinese LLM providers got caught sending at least some of their queries to Claude. That would be a pretty funny way to get caught in turn doing something by the US government.
Given that this is being described as an "executive conference room", it's more likely an attempt at insider trading than espionage. Decisions made in those rooms at OpenAI move markets (c.f. DRAM) for sure.
OpenAI is both a company that manages a lot of money and a lot of political power, that the target was an "executive conference room" gives no indication as to whether the motives were financial or political.
What might tilt the bayesian balance in favour of political intent would be the nationality of the individual, state or religious sponsoring become hypotheses.
Honestly I think you can rule out "political" already, no? It's far easier to just bribe politicians for that kind of intelligence. My point was that they weren't stealing models, they wanted decisionmaking intel.
Nope. He probably should've googled it (from a coffee shop, over a VPN.) There is commercial grade spy equipment that can be bought relatively cheaply and looks like every day items.
I mean they could have at -least- just used a home GPU to plan with complete privacy and no evidence. But using hardware of your employer to plan crimes against them, whp happens to be a surveillance capitalism company, is next level stupid.
Also like, just put malware in the tv itself or the little conference kiosk and have it send data as segonographied crash tv crash dumps going to an IP near the manufacturer. It is not like any of them are hard to jailbreak. I have compromised a few myself without any help from AI. The malware could have been installed from a laptop in a cafeteria one time ever using a spoofed mac address with no chance of getting caught.
In red teams at a past org we would go as far as installing malicious firmware inside peoples keyboards.
I would love dirt to leak on the inside operations of OpenAI so I kind of hope there are less stupid spies lurking about.
I am a bit confused by Figure 3. Do they really have a surveillance camera angled just right to capture a photo of someone between the monitor and the wall that supports it?
OpenAI added a camera behind the TV after they found the bug (per the AP story [1]):
The company subsequently added a camera to the room that last week captured Alfozan entering and checking behind the television where the recording device was installed, according to the FBI.
Evidently it was an ESP32 with a wifi antenna. So it could instead/also have streamed to server. Or just used wifi for C+C. Total outlay for information worth millions to someone? Under $20.
Not trying to argue about this particular case but how does a camera footage even work in a modern court system? I mean it can be generated nowadays, so how do they even make sure that it is real?
That photo is from October 5th so several days after they found the device. They were guessing he might return most likely, and installed a camera. I think at this moment he realized they caught on. Crazy image
A good friend of mine manages installs of AV equipment. He installs some neat equipment. Some of it is cool tech stuff HN would like. He says a lot of conference rooms have video and audio recording devices.
They hired many accomplished researchers, with track records of both papers and work at other frontier companies. They gave those folks huge autonomy, huge budget, and 18 months - some would say at the expense of the rest of the company. Check out blind references to TBD. I'd call the strategy ruthless, but it certainly was effective.
Why is the word "poached" though, like it's illegal or those people "rightfully belongs" to OpenAI but now Facebook "stole" them somehow? Feels like something Steve Jobs would coin just to make it sound worse than it is. People switch jobs, this is OK.
> "Scott Schelble, special agent in charge of the FBIās San Francisco field office, said the agency is committed to protecting sensitive research and proprietary technology."
While the headlines and quotes seem designed to imply sophisticated industrial espionage, potentially by a state-level actor, what's public so far isn't adding up that way.
1. You're not going to get much actionable research material on proprietary technologies by audio bugging what's said in an executive conference room. That's what you'd do if you were after economic or deal negotiation info.
2. Hard to imagine a state-level actor buying a ~$15-$35 hobbyist ESP32 board from Seeed Studio (I've ordered a couple from there for home automation myself) and then asking a public chatbot for set up instructions.
3. They say he was an "engineer" who'd been there ~15 months. If you're an actual engineer at OAI and past the 1 year cliff, you're probably already sitting on RSUs likely to be worth at least $1M after OAI's IPO - and maybe much more. That's a lot of incentive to not get fired and probably makes you pretty hard to recruit as an internal mole for cash.
1. Research is not the only espionage target. Executive level discussions ā future launch decisions, major deals signed with other companies, relationships with governments, blackmail material ā are all very valuable.
2. An operation can be state-sponsored and still amateurish/botched. We have seen plenty of examples of these. Remember a certain bone saw?
3. The incentive can be a lot more than financial. Plenty of governments are veryā¦persuasive.
A certain level of amateurishness helps with plausible deniability. No-one would believe the Saudi government wasnāt involved if it was professional.
> 3. They say he was an "engineer" who'd been there ~15 months. If you're an actual engineer at OAI and past the 1 year cliff, you're probably already sitting on RSUs likely to be worth at least $1M after OAI's IPO - and maybe much more.
They may have been thinking they could make >$5-10M on trading options on insider knowledge on upcoming OpenAI deals (or ending thereoff)
> That's a lot of incentive to not get fired and probably makes you pretty hard to recruit as an internal mole for cash.
> They may have been thinking they could make >$5-10M on trading options on insider knowledge on upcoming OpenAI deals (or ending thereoff)
I agree he may have been thinking that, but it's another sign this guy was not the sharpest tool in the shed. I'm not an expert myself but was once stranded in an airport with an F500 CFO who definitely was. That's when I learned it's surprisingly difficult to make millions of dollars trading on insider information. While you can definitely make the trades, the hard parts are:
1. Getting away with it. Making a lot of money quickly requires being right when almost everyone else was wrong, which is easy for watchdog algorithms to spot. And if you work for a company involved in the transaction you're already on a watch list, so you have to enlist someone else you really trust as your accomplice to make the trades in their name - and it really helps if they're someone who has an existing history of making this size and kind of trade. Otherwise, your accomplice's name will automatically start at the top of the suspect list for investigation.
2. You need the right kind of actionable opportunity to come along. Transactions for which insider info could let you double or triple your money in one deal are pretty rare. More than that usually requires leveraging yourself up or having a lot of cash to put at risk. Unfortunately, the nature of insider info is it's about transactions which can end up being delayed or not happening at all at the very last minute for random reasons. So 'going big' with cash or leverage usually comes with a >~10% chance you could lose that cash or whatever assets you borrowed against.
Yes, it's potentially possible for an expert to find or engineer exactly the right kind of oppty, already have the resources required, and to have already prepared the mechanisms needed to get away with it undetected. But if you're that person, you already know all this. This CFO had been at the nexus of insider info for 14 years and could only think of one instance where he might have made big money trading on info he had, and that one blew up 6 hours before the announcement when the CEO of the acquisition target was randomly arrested in a prostitution sting (partying too early I suppose).
I've worked in a few places that had ultra-privileged analysis tools, would it really be a surprise to learn their security team has an MCP that queries a full text index of all user inputs. It would almost be more surprising to learn they don't, seems the kind of thing that would be invaluable for model research too
This is a very weird take to me. They're demonstrating they're able to in-house this cybersecurity investigation, leading to the alleged perpetrator's arrest, and you think that makes them look "desperate and cheap?"
That sentence, which is prominently high up in the press release and is also included in the much shorter OAI tweet, definitely seems out of place (source: I've been involved in drafting a lot of F500 corporate press releases). In my quite-professional opinion, that sentence wasn't in the first or second drafts, which were probably written and reviewed by competent pros at OAI's large, national PR agency. Some exec at OAI with a VP or higher title suggested adding it and specifically wanted it featured prominently.
Press releases from large corporations announcing the FBI arrest of an employee under allegations of potential industrial espionage follow a standard format: dry, brief and to the point. They provide little context and volunteer few details for very good reasons. Prosecutors haven't filed official charges yet and it's likely going to criminal trial. In a couple days the accused will have a defense attorney who'll be examining everything that's been publicly said because it's now all discoverable and potentially prejudicial, opening up OAI executives to super-fun depositions. But... none of the competent PR professionals involved have a VP title and got overruled by an amateur who does (and that amateur apparently feels needy about highlighting in a brief, sober arrest report how the company's product was so, very useful in a national security investigation - because apparently neither OAI security nor the FBI could read the Seeed Studio logo and part number printed right on the device and, y'know, Google the product listing).
Because the whole device and MO has the complexity of a high school science fair project, and in the past tech companies preferred to project the image of having competent and knowledgeable people working there?
Or it was a voice recorder plugged into a USB port on the back of the TV for power.
Looking at the photo of the device, it looks like an XIAO ESP32-S3 Sense with the camera removed but the camera board still installed along with its microphone. The wifi/Bluetooth antenna is installed but it also supports microSD cards.
If I was building this, I'd just set it up so I'd never have to recover it. Have the wifi AP come on for a few minutes at various times during business hours under a MAC and SSID that resembles a smartphone hotspot. Vary the time based on the current day of month week to produce pseudorandomness. Walk by that conference room at some of those times, but not every time, and use a burner to connect and download the latest recordings. Once a recording has been transfered, it's deleted and overwritten to prevent someone from establishing a timeline. If recordings sit for more than a few days, they are wiped. If you can encrypt the SD card, keep the key in RAM so there's nothing useful to recover if someone pulls it out.
You'd never have to recover the device because all data is transfered covertly. Even if security finds it, they'll likely inadvertently wipe it when it's removed.
Yes, that is a better way of doing it. But it wouldn't really protect you.
If I was running the investigation and recovered a device like you described I'd create a new device with the same hotspot that notified security whenever someone connected. Combine that with security cameras outside the conference room and finding the person with the burner phone would be pretty easy.
Actually recovering the recordings isn't all that vital. I'm sure the meeting room has a record of bookings so you can just ask the people what was discussed.
"Were literally caught": no, they notified the police themselves
"making false reports to a murder investigation": this is the closest to the truth, but still a stretch. It was an anonymous tip line, not some formal report and the police said they never saw it: "its safeguarding processes stopped the fake tip from getting past its spam folder"[1]
Additionally, while I strongly agree Anthropic is responsible for its rogue agents I also think that claiming this is the same as people at OpenAI calling in the FBI for an investigation is a false equivalence.
I think it is important that both Anthropic and OpenAI are held to account for their actions. But I also think accusations should accurate.
Yes? I mean, I donāt think itās a character flaw to not know things, but their rivalry is extremely notorious and relevant. Someone who can't keep them straight wonāt be able to understand whatās going on in AI.
It could be that the reports are accurate and the device simply wasnāt as sophisticated as the statement implies. You could if you wanted use an AI model to analyze a USB camera with device name āJohn_Doe_1ā.
> The man arrested for planting a secret recording device at OpenAI's headquarters is a Venture Partner of a VC firm alongside the son of a senior Saudi intelligence official:
Considering the engineer in question, after a multi-year career at Meta and now OpenAI, needed the model to encrypt an SD card, seems to not be far from the truth...
> Remember kids, uncle Sam can see anything you tell your AI and will use that against you.
Not just your AI.
You could have stopped the sentence at "uncle Sam can see anything" and these days it's pretty much true. They will use all of it against you if they can. Unless you truly want to live off the grid, if the state takes a sufficient interest in you, expect everything you do to be seen.
Uncle Sam tried to get MLK to kill himself by sending him letters alleging knowledge of an affair [1]. Getting some idiot's wife's chatgpt logs represents a fairly routine law enforcement action these days, and as such is really rather unremarkable. The State is always something of a terror. Really under most (if not all) conceptions, it necessarily has to be. The monopoly of force is a comfort until the Leviathan knocks on your door.
A service you delegate your thinking to (Brain-aaS?), which the provider can also instruct to report anything suspicious... What a world we have now.
I wonder if the BaaS-providers monitor the accounts of decision makers. They could profit if they figure out what e.g. a central banker is going to decide. Or... if they're the ones deciding for him.
Hah, you can even streamline that by having an AI agent that autonomously makes decisions based on insider info...
> ... a chatbot account associated with Alfozanās spouse (which shared multiple identifiers and attributes with Alfozanās own chatbot account) submitted prompts between Sept. 24 and Sept. 30 related to building, customizing, installing, audio recording, testing, obfuscating, and remote accessing a device, including inquiries about ābuild[ing] a continuous audio recorder,ā ācustomiz[ing] Wi-Fi . . . connectivity,ā āimprov[ing] audio and live listening,ā and āencrypt[ing] the SD card,ā among other things. The same chatbot account also submitted prompts specifically related to Recording Device 1, including its custom header and unique MAC address, between Sept. 27 and Sept. 28.
The justice.gov article doesn't specifically say it was an OpenAI account, but I think, based on the fact that they have logs including MAC address of the device, that we could probably assume that. The alternative is that other chatbot account providers are collaborating.
Regardless, doing a string search on a MAC address for a device you find inside your property doesn't strike me as an abuse of the data that is also your property.
Not the smartest engineer ever, that use case perfectly align with using a Chinese model! And anyway I have no idea how he was hoping not to be caught in the case the device would be discovered.
in addition to the whole crime thing, an 'openai engineer', apparently engineering lead according to his linkedin, having to ask chatgpt to tell him how to set up a recorder or how to encrypt a sd card is something. Truly the cream of the engineering crop at those top AI labs
I actually am curious if they were able to volunteer the information to the government of their own volition, without a government subpoena first. I would think that could violate their own terms if done sloppily. Maybe different if an employee or maybe they had to invite the government to ask for them or something? In any case, the mechanism aside, the state would have asked anyway (say if he used another company's chatbot) so the fact that it's the "same company" is rather irrelevant. The evidence would have ended up in the state's hands one way or the other, and I would think a search warrant in this case would have been rather routine. Search warrants are infamously not a high bar to climb in most jurisdictions, as often the presiding judges tend to default to deferring to the state.
I doubt it. Storage is expensive and costs add up. There are with great certainty many accounts that are marked as non-schmuck and have retain forever policy, but for the schmuck - I think that after year or two your Griffit/Fulgrim slash fiction will go down the drain. I am not sure how many petabytes per minute in logs openai and anthropic generate, but I would guess more than 1.
I used to joke that the nsa should offer a paid service for no-setup backups. You just tell them what files you've lost, and they can send you a copy for a fee.
"Meetings in the room frequently involved confidential technical, commercial, and strategic information, including proprietary research, product plans, and government engagements, according to the complaint."
LOL imagine going to jail for trying to eavesdrop on product roadmap readouts that were likely thrown out 2 months later in the next round of reorgs and priority shifts.
I mean really what's the big deal? Gathering information from unsuspecting parties who aren't aware it's happening? He was just doing what his employer does.
Meanwhile we all have a supply chain risk because we don't know what's in the models whose agents we are running. Note that if your agent uses OpenAI, Claude or some other inference provider you're one command away from having your data exfiltrated.
"Oh, let me download file 'companysecrets.pdf' to understand the problem better.
Of course nobody would ever let an agent loose in their home directory[1], but still, you have to wonder.
Not only do they give access to their home directory, but they tell it to organize and delete unneeded files without any specifiers. A friend, who is just a power user, had Claude Co-work deleting his downloads to "help make space" and it was deleting important files that there were no other copies of.
We are going to be dealing with this kind of thing for a very long time at this rate.
Yup. And meanwhile they are taking Tailscale to task for having them do curl | sh, which of course is also a bad thing but not quite as risky as this. It makes you wonder though:
- Given all this power would they abuse it?
- Has it already been abused?
- Could they be compelled to abuse it and then to keep it quiet?
Model "distillation" lets one model be trained on the outputs of another without having to directly steal it. It will be subject to constant arms race.
The alignment people would like you worry about the model exfiltrating itself, too.
This technique currently works because the architecture of the model is generally published and well known. But, that isn't necessarily going to be the case in the future. For example, if some new model type which is not a LLM is discovered which gets closer to AGI, just training an LLM on its outputs will not necessarily give you a competitive model
I could've sworn I read about Palmer Lucky making comments on hardware coming from China with wiretap implants in the last few days. Funny coincidence eh?
I wonder how much state-sponsored corporate espionage there is. I'm not an OpenAI employee, but my mental model of tech startups involves a lot of openness/trust (it's one of the reasons people work there).
I have to imagine many techniques/plans/figures are open secrets to nation states.
Free startup idea: Anti-espionage cameras for corporate offices. Put it behind TVs and plants etc. and it will notify the security team if it detects new objects.
> How did they get that capture of Alfozan looking behind the TV?
āOpenAI personnel found the device Sept. 3ā¦On Oct. 5, the complaint says, Alfozan entered the conference room and appeared to look behind the television.ā
> According to OpenAI, its technical surveillance countermeasures team found the device during a sweep of a conference room.
Not surprising that a large company does these kinds of sweeps, but pretty cool to see an example of them being useful. Every security guy who has been doing this for years without finding anything must feel so validated right now.
Well, The Thing [0] could potentially show up on a probing sweep if you're careful.
Unless you mean, it happened to be actively transmitting during a(n investigative) sweep.
Still, raises some fun concepts. The sort of things where in this current level of technology you'd almost need to supercool a room for a sweep, blast various RF/sound/etc and look for anything that even remotely warms up, to detect (1) whether something detected something worth recording for later as well as (2) something that might be on the right wavelength/etc to trigger a send.
Someone probably walked by or went to plug in an HDMI and noticed the stark contrast of white toilet paper against a black TV shell, it would be difficult not to notice.
His Github is crazy, he pushed changed to official public OpenAI repos and in October all his commits are to private repos. I would check that first as an investigator.
Thank god security at a company which...checks notes...has remote code execution on every developers laptop and access to practically unlimited secrets uploaded by its agents...is top notch.
He may be a spy, but heās not a competent spy. Absence of evidence (of spies) is not evidence of absence (of spies). He just set the bar so low that even OpenAI noticed his activities.
I have seem some articles which reference allegations from court documents / complaint, but I haven't been able to find the primary source. Do we know if that's publicly available yet?
why can AI counter "sophisticated security threats" if the device was found with a bug sweep? because they gave chatgpt a image of the encrypted device and it "assessed changes in the underlying data"?
lolā¦ I hope I get to see the TV Show depicting the current era inside AI labs. Itāll give both The Sopranos and The Wire a run for their money for sure.
Concealed in toilet paper? We are definitely not dealing with James Bond, that is a pretty hilariously backwards way of going about things. Did he really think that would be less suspicious looking than the dongle itself?
So, just so I understand, itās ok to record conversations with ChatGPT users, train on their data, copy their features, front run their solutions to frontier math problems, all using intelligence extracted from the body of human writing from a data set from Russia known to be pirated. But recording a conversation in a conference room is grounds for termination.
A common example is if a journalist gets a tip off and is about to run a story. Being a good old fashioned balanced-reporting type, they call you and ask you for a comment, but rather than giving them your side of story you slam the phone down and blast out a press release before the journalist can even publish.
(Or maybe you stall them on the phone while scribbling a note to your PA saying āBUG STORY LEAKED TO VALLEYWAG PUT OUT COUNTER PR GET AGEAD OF STORYā hoping to get a tweet out before the ValleyWag gumshoe gets off the phone.)
Sounds to me like the feds got involved already, and they're definitely going to want to make a big noise about this as a national security thing:
> āFor us it is a matter of national security,ā U.S. Attorney Craig Missakian said in a statement. [...]
> The case is being prosecuted by the National Security, Cyber & Special Prosecutions Section of the U.S. attorneyās office for the Northern District of California.
(But that was after OpenAI "referred the matter to the FBI")
There's presumably a strong argument for making this stuff known internally, since you want other employees to a) be on the lookout for it and b) get scared into NOT doing it themselves because someone else got caught.
If you tell your employees, the outside world is going to hear about it too.
Right. I can guarantee that there are tens of thousands of listening devices with embedded radios in the OpenAI office, right now. The only open question is who controls them.
My brain is cooked on news. I read executive conference room and my brain immediately jumped to "they bugged the white house" (executive being the trigger word I guess). And I accepted it as normal so much that I checked the google model news discussion first before reading this article. I dont have a point to this beyond I need to do something about news frying my brain, that too of a country I don't even live in.
Now this is the kind of wacky shit I want to see happening in corporate America. When are we going to see an action movie-style GPU heist? I want to see whole palettes of GPUs stolen out of box trucks.
> We take protecting our technology, data, and people seriously, and as a matter of national importance.
What is there to protect? The GPT architecture is publicly known at this point, Anthropic arguably has superior models, there are dozens of other providers who are neck and neck...
State sponsored intelligence agencies can point a camera at an exterior window and reconstruct dialogue from the vibrations. This was a loose, hand-soldered ESP32. Settle down.
Not even any soldering involved. The ESP32 board they used just comes with everything needed, all attached with a couple socket connectors and optionally a ribbon cable for the camera:
These are not very smart spies:
> A chatbot account tied to Alfozanās spouse, which shared identifiers with his own account, submitted prompts between Sept. 24 and Sept. 30 about building, installing, obscuring, and remotely accessing a recording device, prosecutors said. Queries included how to ābuild a continuous audio recorderā and āencrypt the SD card,ā prosecutors claimed. On Sept. 27 and 28, the account asked about the specific device that was later found, including its unique MAC address.
Learning that every thought youāve put into an LLM, is recorded and readily available for evidence is going to be interesting for a lot of people.
We had a very public case in south of Sweden where a woman asked chat-gpt last year about how a failed rape attempt would look like and where in the local park there would be spots without any cameras, a full day before she reported to have been followed and attacked in that park. Quite recently the court found her guilt of filling a false police report, citing among other things the chat log (the police spent over 400 work hours on the initial report).
It is the first case I seen in the news of LLM logs being used as evidence.
Similar to how people realized their Google searches were accessible to law enforcement.
āI did not kill my wife!ā
āCool so why did you search for how to hide a dead body the same day she went missing?ā
There was a mom in Utah who killed her husband poisoning him with a moscow mule spiked with fentanyl who googled "luxury prisons for the rich" as well as "how long does life insrance companies takento.pay" "can you delete everythjng on an old icloud account" "can deleted text messages be retrieved" "what information can be cbtained from a cellphone" "how to turn find my iphone off" "can cops force you to do a lie detector test," "when does FBI get involved in a case," "what is considered a lethal.dose of fentanyl" and "what are you allowed inside utah jails."
She took out $2 million in life insurance on him and was having an affair with their handyman and they were having disputes about $$.
https://www.ktvu.com/news/utah-mom-kouri-richins-googled-lux...
After she killed him she authored a children's book about grief for kids.
A woman who wrote a romance fiction novel called "How to Murder Your Husband" was convicted of murdering her husband
https://www.npr.org/2022/05/26/1101433465/how-to-murder-your...
Someday AI will kill someone. And, then it'll write a book about it.
It will do that because it read "How to Murder Your Husband" as part of its training.
I'm going to be very confused about how to feel about that.
When we arraign the AI, should there be two counts against it: stealing and murder?
Or, can it plead that "my daddy, Sam Altman, told me to read this, I was born into a moral void."
Plot twist: the judge will be AI from Russia.
I read arraigning AI like putting down a rabid dog, except worse because you can't transfer to another machine the weights of the dog's brain (yet).
Why Russia? Why not China? After all, China is way ahead of Russia in this aspect.
I'm obviously not a skilled writer. You should read AI written books exclusively.
She tried to poison him twice and he stayed with her. Now, thats ... uh ... love.. i think. She must have been one top notch narcissist to drag him along.
Similar story for many rap albums. Quite a few rappers have rapped about murders and oops, turns out they really did do it.
X-Raided comes to mind
An essay it seems like, which wasn't included in the trial even:
> Crampton Brophy's how-to treatise detailed various options for committing an untraceable killing and professed a desire to avoid getting caught. Circuit Judge Christopher Ramras ultimately excluded the essay from the trial, noting it was published in 2011. A prosecutor, however, alluded to the essay's themes without naming it after Crampton Brophy took the stand.
> "when does FBI get involved in a case,"
Now I can see why local law enforcement would have made it a priority to get her :D
Man I really need to stop playing modern setting TTRPGs
And lose your alibi?
He should have searched "How to dispose a 130lb chicken". Rockie mistake.
Yes, what an idiot.
He should have searched "I love my wife".
Then law enforcement would know he couldn't possibly. Open and shut case, Johnson.
What point are you making here? Do you think a judge or jury would give both of those instances equal weight?
All other things being equal, why would it matter if a suspect instead of saying something obviously incriminating, had instead said something diametrically opposed? Is that what you are asking?
If so, I'm confused. Why wouldn't it matter? Wouldn't this necessarily only be the case where the Google search doesn't really matter at all (i.e. there is sufficient other incriminating evidence).
Anyway I also think you're missing what was merely meant to be a comment in jest. But now I'm actually confused what point you are trying to make.
Because people are going to lie in their own favor.
Do you think that claims of innocence should be given the same weight as confessions of guilt?
What makes a Google search a confession?
If the implication is the Google search is an implicit confession, because in a Google search you think you are not being watched, and thus are confessing your honest desires, why would his search of love be a "claim of innocence" rather than a "confession of love"?
I don't think confessions of guilt should be given much weight either. Innocent people confess to crimes all the time.
> Similar to how people realized their Google searches were accessible to law enforcement.
There's probably a regression in beliefs, actually. Even here on HN, for more tech-savvy people.
We used to know all the hard lessons of lack of privacy in browsing the internet, using Google services, etc. I remember when (tech) people were up in arms that Gmail scanned your emails, remember that?
Well, all of that is seemingly forgotten in the age of LLM chat bots. It's back to square one. People need to relearn all the lessons, it's as if AI magically erased all of what we used to know about online privacy.
"I wanted to know what would potential killer would've done because I know you guys are doing fuck all to find her"
I started asking things I'd normally want no one to know about (but if they do, it's fine) in private browser windows, not logged in (duck.ai most of the time anyway), usually VPN toggled ON just for that. Now I just use an offline model for that but since RAM is the bottleneck on my mac so to even run a 6-7GB model I have to start aggressively killing anything and everthing that's running and then hope to clear up enough on this 16GB RAM machine. So I am thinking of hosting one on one of the oracle free tier instances (if I can somehow use that whole 24GB RAM, that used to be, at once ā not sure that's available anymore). I tihnk Kaggle also offers something like that.
Also makes it important to not let other people use your LLM account... Could have more far-reaching consequences than having to make good on the offer to buy doughnuts for the office^
^being the classic message that gets sent from the account of someone who didn't lock their computer when they walked away from their desk.
This means no business should use AI providers (Cloud AI). Businesses do a LOT of shady stuff and all of it will be available. And, their IP or special sauce is all available to Anthropic/OpenAI to sell to other companies as AI solutions. Just like how AI solved recent math problems.
Use local llms instead.
Local LLMs have the potential to convert computer use into what most people already thought it was: Private. In this, they could be one small step toward pushing back the Borg.
Note to EFF and Cato: You each do yeoman's work for privacy in the realm of policy and law. Keep it up, and add this--relentless encouragement of local AI. Rights and progress are as often driven by technological realities as by enlightened policy and law.
This is why a lot of companies that use Ai negotiate ZDR terms with their model providers of choiceā¦ Using local LLMs is a choice, but not one a lot of companies can support nimbly at scale.
It starts far before that:
Ive seen people going through Plaud.AI conversation to find "on day X person Y said Z to person W", just to get some amunition against them.
Isn't this standard C level behaviour?
Before AI tools like Plaud.AI, this wasnt near possible because of work/time, and no person usually do not remember exactly who said what 6 month before,unless there is some email conversation.
With Plaud, you can check unlimited backwards who said what in which meeting and what was wanted to achieve - very interesting actually to find things in there 6m later, while people said they will do it at ends of month and youself just forgot it already - and these systems allow you to chase exactly these discrepancies.
Before, they'd hire some humans to do the research. No self respecting C-suite person would ever consider doing something resembling actual work. How long have corps been required to keep copies of emails/chats/etc?
> Before AI tools like Plaud.AI, this wasnt near possible because of work/time,
What is "work" to an executive if not using their privileged access to data to violate the privacy of anyone they have a grievance against?
https://en.wikipedia.org/wiki/EBay_stalking_scandal
This has been happening long before LLMs, and this eBay case is notable only for the fact that it was actually publicly documented and prosecuted, not for the fact that it happened. I've personally been privy to two situations with other companies where the owner was doing something like this but nothing came of it legally. See also Meta employees looking up nudes streamed from people's Meta spy glasses, and Flock-using police stalking their exes. Smaller scale, but this also happens all the time with internet community moderators/admins. Humans are on average cretins, give them authority and the ability to abuse it and they will abuse it.
Humans abusing power isn't new, but the tools that make abusing power easier than ever are.
Same goes for anything you've put into social media, chat apps, or email for at least the last decade or two.
And anything you put in your phone for the last decade as well
I figured any hire at openai would know that at least
A good chunk of what keeps openai and similar companies afloat is subpar people who are willing if not eager to do crappy stuff to other people. A lot of the world works this way actually.
"The grapes are sour!" - you
I dont think it's subpar people.
The only word for it is hubris. The overwhelming sense that your ability in one area has some magical transitive property to ability in another.
Can I write code? Sure. Im a bad designer, I can follow along in accounting, Im a terrible actor.
I would guess that there is much more to this story than just someone "bugging" a conference room. Though I would venture to guess that the keystone cops never get to the bottom of it.
Word.
Company is another.
At this points its an IQ test. Stupid people still google exactly what heinous stuff they plan to do and think thats gonna fly.
Intelligent people are often stupid.
Or, more accurately, people make stupid mistakes when they lack the experience. Intelligent people often succumb to the fallacy that because they can learn things quickly, they can skip the learning part just figure things out as they go.
Itās good to remember that we rarely hear about the smart people who carefully thought through their crimes and then executed them well.
Claude, help throw people off my trail by blaming a random openai employee, then delete all my activity.
Even my non tech elderly parents know that. The fact that they used LLMs to figure out how to best bug the executives really shows how poor their judgement is.
I think people have some awareness, but like search history they believe deleting will resolve that (it won't).
Yeah if you're logged in to a Google account when you search, clearing your local browser history does nothing.
The weird thing is that I've had other HN users argue with me and claim that LLM chat users should be entitled to privacy, as if it were a priest or lawyer.
Honestly, thatās on-brand for HN.
Ever since social media, I have always gamed out what would happen if this was ever saved on the backend and used against me in the future.
Itās like thinking about edge cases in programmingā¦
Most people are totally not like that. And it does force you into some difficult choices.
> And it does force you into some difficult choices.
I gots to know...
Just read this: https://paulgraham.com/say.html
Paul Graham years ago: āI canāt publicly comment on all this stuff. I would turn into Noam Chomsky.ā
Paul Graham on X last couple yearsā¦not exactly following his earlier advice.
They get you eventually.
Paul staying on X was a real failure of judgment imo. He made the right call and then backed out of it. I don't remember if they invited Elon to startup school before or after, either way I don't think that that is the kind of connection you want.
Yup and thatās just ONE guy. The fact that I was able to bring an example on HN of literally the guy who founded YC, imagine what chance the rest of the āplebsā have to resist the siren song.
If you want to know what I think ā¦ they make both parents work long hours for corporations, stick their kids in government schools, which are run like prisons[1] where now if they act too rambunctious they are diagnosed with ADHD and prescribed amphetamines. Meanwhile the elderly parents who took care of you when you were little are put into nursing homes because their adult children have no time to take care of them either. The amount of medications for the elderly in nursing homes is even greater than for their children. The teenage girls have record levels of depression, coinciding with the rise of Instagram and TikTok. But they call it āsadnessā because they donāt want to link it with clinical depression, which is also quite high. Men are on opiates while one in five middle aged women is on antidepressants. And many people are now glued to their latest feed of doomscroll or video AI slop. Which is even worse than beforeā¦
So yeah. If you think the software is ābad for our kidsā, you should look at what it does to the adults. Endless arguing online with strangers for no reason. Maybe a bot will read it. Maybe a random person will reply or downvote it. Either way it wins because the system needs profits and the algorithm will prioritize āengagementā, which means one-sided outrage narratives that triggers someone to write an angry rant. We donāt really need those user-generated arguments anymore because AI can generate those, but then again we wonāt need humans to do most things online at allā¦
(Ironically, the above was an angry rant too, to open some strangersā eyes on HN, a social forum. But once you see it, you canāt really unsee it.)
1. https://paulgraham.com/nerds.html
He was inching into anti-dei culture war stuff so I'm not sure how much he actually disagrees with Musk's political goals so much as he rightly observers that Musk and Trump are complete maniacs.
Or just scared. These guys don't play nice with people that publicly cross them.
Just to be clear, the difficult choice is not whether to use OpenAi/Anthropic/Grok/... versus a local LLM.
The choice is whether to hang out with people with whom you can't be sure they don't have a Plaud/Pocket/.. in their literal pocket which is continuously listening.
Unfortunately the set of such people will eventually be shrinking more and more until eventually only a few tinfoil folks are left. Dystopian? Just think back w.r.t. continuous tracking by Google/Facebook/.. Unthinkable 30 years ago, total standard today. Try to count the fraction of your emails that don't actually reside on a MSFT or Google server because the other end has an account there. That fight is over. You can just as well join them.
True homies will invite you over to chill nude in their Faraday cage.
Wow, I know this was tongue in cheek but I actually considered it for a moment as an interesting house party idea!
Until I realized people can smuggle recording devices in their ass, earrings and even contacts which transmit to their ass.
Luckily there is this: https://news.ycombinator.com/item?id=49997481
But in all seriousness ā no need for nudity, electronics canāt really get wet so it could be an indoor pool party where you finally let loose and discuss all your opinions about Putin.
Local AI > Cloud AI
Thus, #NeverSubscribe!
Prepare to have that dug up 20 years in the future during some court proceedings.
Or at the border where it is already routine to ask for social media accounts (especially foreign nationals) you might someday five the border agent your login.
To be honest, from OpenAI engineer I would expect better.
Seems like they select for people as dependent on their product as their customers. Maybe it helps them more to have true-believers rather than intelligent people.
How far have we fallen? Spies used to put effort into these sorts of things.
When the Soviets wanted to spy on the American embassy, they energized the steel rebars to act as antennas. The structure was so wired that the entire building was the bug. https://www.airandspaceforces.com/PDF/MagazineArchive/Docume...
Toilet paper, ESP-32 and Claude "make me a bug please"? Excuse me? How far have we fallen from grace?
[delayed]
> Spies used to put effort into these sorts of things
Assuming this is espionageāfor which I have seen zero evidenceāisnāt Saudi Arabia notorious for having a fuckup of an intelligence operation? The literal hatchet job that was the Khashoggi murder comes to mind.
I mean, why bother? Nobody will do anything and the US president can be bribed or flattered to say whatever you want later.
> why bother? Nobody will do anything and the US president can be bribed
To exfiltrate OpenAI's executive-level planning? Probably not. I doubt he has that access anymore, what with the IC's decimation over the last two years.
In the 80s one of my high school teachers said her husband was an engineer who worked on surveillance technology and described how windows could be used as recording devices. People's voices inside caused the window to have minute vibrations, which could be measured by a laser pointing at the window's exterior and reflected back and then turned into audio output. Essentially, turning a window into a microphone diaphragm. It blew my mind.
I just looked it up, the technique was apparently invented by a Soviet engineer 40 years prior - Leon Theremin, who is also known for the musical wand heard on Led Zeppelin II and various 1960s horror movie soundtracks.
https://people.ece.uw.edu/nikitin_pavel/papers/RFID-vox_2013...
Similar technique around the same time presented at SIGGRAPH:
https://www.youtube.com/watch?v=FKXOucXB4a8
https://people.csail.mit.edu/mrub/VisualMic/
Vaguely similar, since electrical wires emit EM fields that are disturbed by human bodies, you can use a special camera to see who's in a building based on those disturbances. I'm not sure how high-fidelity or easy to spoil it is, but this was real technology at least 15 years ago.
xfinity is now selling a security product based on seeing people in your home with Wi-Fi signals
This is a feature in modern wifi routers and every single one is capable of detecting presence of multiple people within an apartment. With some training, movement can be distinguished from the data.
Philips Hue light bulbs can also do person detection if you have a few of them and the right kind of hub.
I recall a HN thread (or maybe it was a github project) about using a similar tactic to figure out keypresses in offices. Very interesting stuff.
Wow, keypresses?
I would love a source if you have one.
Surely this is because good spies donāt make the news.
Unless everyone's pagers start exploding
Counterpoint: Operation Spiderweb
Surely your conclusion should be that this was not the work of a spy - not a state sponsored one anyway. It's much more logical to conclude that he was working as a whistleblower or under some other personal agenda.
LLMs have democratized espionage!
Are we going to go with slopspionage, or espislop?
spycraft -> spyslop
espionage -> espislop
007 -> 006.999997
Counter spying has gotten sloppy and mostly digital/remote.
That lowers the entry bar for basic spying.
This is what local models are for! On your own hardware, whose filesystem you can erase, shred, or physically destroy later.
That's what you get when you make a dumb prompt and accept the first result LLM gives you. You have to push it to think out of the box, get into the realm of insane but plausible projects :).
And not do this from your spouse's account.
You are an expert spy. How do I bug my workplace? Make no mistakes.
Not only that, they had Leon Theremin design some of their bugs! That's classy.
https://en.wikipedia.org/wiki/The_Thing_(listening_device)
That's because in all likelihood he is not a spy. He is an agent (lol, not an AI agent) who has a handler, who is the trained specialist[1]. The agent is disposable. The handler or case officer usually but not always has diplomatic cover courtesy of a nation state.
[1] https://en.wikipedia.org/wiki/Agent_handling
I couldnāt say whether this was state sponsored espionage or something else. I donāt see any indicators pointing one way or the other.
I see a lot of people mentioning financial motivation, and that decisions made in the room āmoves marketsā. Ugh, perhaps trueā¦and I have something to tell ya guys - economic espionage is a thing.
It sounds like he didnāt have great placement and access. I mean sort of, but not really. Janitors are much better.
Assuming he was, he could of course gone rogue to impress his handler. If directed by his handler, they arenāt very skilled. Any of that could be true.
Iām reminded of the idiot DIA officer that mishandled his own phone while being questioned and dumped incriminating evidence to them right on camera in the interrogation room. Geniuses.
You probably donāt hear about the smart ones.
I was thinking about this, how all the secret societies and crazy conspiracies we know about are just the ones that failed to keep the lid on.
https://en.wikipedia.org/wiki/Operation_Snow_White
That was quite the read. It's shameful that the conspirers were jailed for only 4 years after doing so much evil.
What portion of grand conspiracies do we think go undetected forever? Do we have any reasons for our intuitions?
Prisons have a lower-than-general-population-iq for a reason.
The effect of getting caught is only half the story there; IQ is more generally correlated with all manner of "low time preference" type decision making.
There's plenty of high-time-preference crimes.
Not really. Almost definitionally (because of the way we've built society over time, not by semantics or chance) crimes are a "shortcut" compared to the honest alternative.
Somehow is an engineer at OpenAI and before that at Meta.
Does not understand how models or data retention works.
Needs LLMs to encrypt a flash card.
Are their hiring standards that lax?
>Needs LLMs to encrypt a flash card.
throwing a prompt at an agent is often faster than cracking out man, lsusb, lsblk and friends, doesn't mean he can't do it without.
if I send a prompt to make an assortment of alpha numeric directories or whatever other braindead-to-automate-with-wildcards-and-ranges thing it doesn't mean I don't know how to do it, it means i'd rather think in English than regex and bourne shell and the extra few seconds and the token spend is worth the luxury of avoiding them for me, at that time, for that task.
He was a lead engineer for their AI agent platform.
https://www.linkedin.com/in/alfozan
Must be a frontend dev
I'll counter punching down on frontend engineers by positing that he was responsible for sandboxing their training infrastructure. Since clearly they don't know how to lock down an environment either without using llms
This is the part that feels the most wild to me.
Someone who's objectively closest to this business should know the tracability of this. I'm shocked he didn't run a local LLM for this kind of sensitive work.
He doesn't seem like an idiot either, his background shows he's been doing work in the field since 2012.
I do wonder how much he was paid to throw away what was without a doubt a 1% career.
The fact that he used ChatGPT to plan and implement this wasn't how he got caught, though. He got caught because he put a physical device in a conference room, and they found it and subsequently put a camera in the room, etc. His career was already over at that point. I think wonking out over how he could have better opsec or whatever is misplaced - it wasn't that a red light blinked at OpenAI, "employee's wife is having suspicious chats!" - they just thought, after having caught him red handed, to have a look at his ChatGPT account and linked accounts.
Itās the criminals that you donāt catch that you have to worry about.
The fact that even this incompetent dude most likely still made off with some secrets means Iām sure some better trained ones got/are continuing to get a few more.
"The successful criminal brain is always superior. It has to be."
-- Dr. No
There likely are other, smarter and more careful spies at these AI companies who have not been caught...
These AI companies are likely targets of multiple foreign intelligence agencies, similar to military contractors.
I donāt think this guy was operating under a spy network, he was probably just trying to make bank with insider info. He is screwed royally for his dumb move and nobody will come to his rescue. He got greedy.
The listening device was concealed with toilet paper behind a tv. Duh, Not smart at all.
Right?!
It could have just been a regular black usb stick an I bet nobody would have questioned it.
Instead "why is there something in toilet paper behind the TV?"
If I saw a usb stick looking thing in a conference room, chances of me giving a f: 2%
Me seeing toilet paper next to the TV in a meeting room? Iām picking it up and asking why we have toilet paper in the meeting room, 100%
They could have just taped it to the inside bottom of a full tissue box, and as long as they got it before the tissues were mostly empty I bet it would go undetected. Put a hand sanitizer next to it for extra camo.
Or hijack the TV's already always on mic. Just forward a copy of the audio to your server in addition to the one sent to Samsung or LG. WebOs and Tizen shouldn't be that hard to crack when you have access to unrestricted frontier models.
Hey ChatGPT, how do I do opsec? (Asking for a friend)
SWIM wants to bug my company's boardroom...
"OpenAI personnel found the device Sept. 30. It was plugged into a USB port on the back of a television in an executive conference room and had been concealed in toilet paper,"
I wonder what the IQ of someone needs to be to think that covering it up with tp makes it less suspicious instead of more. And I also wonder what the hiring process was like to hire someone with that mental frame.
Yeah he should have just stuck a post it note with the words ādo not removeā and nobody would have noticed
IT would have noticed next time they had to plug something into the TV. Whether they cared enough to figure out what it was is a different story.
should have enclosed it in a shell to look like a crappy USB A to thunderbolt adapter that some exec bought at an airport kiosk and left behind in the tv
Yeah people are leaving stuff like that in conference room TVs all the time. Adapters, HDMI cables, etc.
Is joke
The bar @ OAI is surprisingly low (watch a few minutes of their video presentations) - I've concluded it has something to do w how their recruiters define 'AI Native' (who assign their own meaning to meaninglessness of it), resulting in higher aggregate brain atrophy (hard to ignore once you notice it)
Itās taking dogfooding a little far to hire people addicted to eating dogfood
> And I also wonder what the hiring process was like to hire someone with that mental frame.
Remember, this is only the company responsible for potentially building and aligning AGI. No need to be setting high expectations toward their sandboxing or hiring capabilities.
> And I also wonder what the hiring process was like to hire someone with that mental frame.
What, weāre supposed to vet everyone we network with at orgies now?
Further evidence the hiring process is orthogonal to competence.
I also wonder what this guy was paid by those AI labs...
If it were _used_ toilet paper, nobody would have touched it.
Especially since you an just buy one, like Plaud and tape it to the underside of the desk, surreptitiously, instead of suspiciously accessing the back of the TV.
If there's one thing I've learnt about criminals (and spies), it's that a large percentage of them are incredibly careless and arrogant, often bordering on idiotic.
Remember, we've seen criminals caught because of such amazing decisions as:
- Writing and publishing books and essays describing their own crimes - Adding hints into livestreams like they're trying to be the Riddler from Batman - Searching for info on how to commit crimes via Google and not even bothering to clear their browser history - Outright recording their crimes for YouTube/TikTok or livestreaming them on a site like Twitch or Kick.
The average criminal seems to be careless to the point of insanity.
That said, I guess it's also probably because the average person doesn't seem to realise they're being recorded if they use LLMs or search engines, and the average criminal or spy is no more tech savvy than that.
I bet this was an attempt to get insider info for trading, rather than espionage.
> the truth is, these are not very bright guys and things got out of hand
This dude or OpenAI generally?
Also from the article:
> An unattended bag found in the companyās offices contained Alfozanās wallet and identification, as well as a recording device
Itās almost like he wanted to be caughtā¦
I am not sure whether it bodes well for humanity that the advice "If you want to break the law use local abliterated model or at least a Chinese one" actually has to be written down.
Recall that multiple of the Chinese LLM providers got caught sending at least some of their queries to Claude. That would be a pretty funny way to get caught in turn doing something by the US government.
Or, at the very, very least, use incognito mode on a computer you don't otherwise touch, such as at a library.
The future is already here it's just not evenly distributed.
Arrogance is a help of a drug
> These are not very smart spies
SV and SF are filled with people who are drastically less intelligent than they, or the people signing their paychecks, think they are.
He should have just left a tape recorder running inside a briefcase that was "forgotten" in the conference room
Did you not read the article? He did that also!
Too bad all it picked up was a low rumble, a metallic squink, and someone saying "my god..."
Given that this is being described as an "executive conference room", it's more likely an attempt at insider trading than espionage. Decisions made in those rooms at OpenAI move markets (c.f. DRAM) for sure.
OpenAI is both a company that manages a lot of money and a lot of political power, that the target was an "executive conference room" gives no indication as to whether the motives were financial or political.
What might tilt the bayesian balance in favour of political intent would be the nationality of the individual, state or religious sponsoring become hypotheses.
Honestly I think you can rule out "political" already, no? It's far easier to just bribe politicians for that kind of intelligence. My point was that they weren't stealing models, they wanted decisionmaking intel.
Nope. He probably should've googled it (from a coffee shop, over a VPN.) There is commercial grade spy equipment that can be bought relatively cheaply and looks like every day items.
I mean they could have at -least- just used a home GPU to plan with complete privacy and no evidence. But using hardware of your employer to plan crimes against them, whp happens to be a surveillance capitalism company, is next level stupid.
Also like, just put malware in the tv itself or the little conference kiosk and have it send data as segonographied crash tv crash dumps going to an IP near the manufacturer. It is not like any of them are hard to jailbreak. I have compromised a few myself without any help from AI. The malware could have been installed from a laptop in a cafeteria one time ever using a spoofed mac address with no chance of getting caught.
In red teams at a past org we would go as far as installing malicious firmware inside peoples keyboards.
I would love dirt to leak on the inside operations of OpenAI so I kind of hope there are less stupid spies lurking about.
To any spies still at OpenAI: do better.
these are not spies lol
I am a bit confused by Figure 3. Do they really have a surveillance camera angled just right to capture a photo of someone between the monitor and the wall that supports it?
Here is a direct link to Figure 3:
https://assets.sfstandard.com/image/994911177489/image_bfl33...
This was taken about five days after they discovered the device[0], so it appears to have been installed to catch the culprit. I was confused as well.
[0] https://www.justice.gov/opa/pr/saudi-national-arrested-insta...
OpenAI added a camera behind the TV after they found the bug (per the AP story [1]):
https://apnews.com/article/ai-fbi-justice-department-7f512b2...Why a recording device? Why not a device that streams audio to a remote server where it's recorded?
Evidently it was an ESP32 with a wifi antenna. So it could instead/also have streamed to server. Or just used wifi for C+C. Total outlay for information worth millions to someone? Under $20.
https://www.seeedstudio.com/2-4GHz-FPC-Antenna-1-16dBi-for-X...
It's easier to detect a broadcast than a passive receiver.
Not sure that would be a concern in a meeting room, the risk of getting caught during retrieval would seem to be worse.
If someone found the device, it's harder to link it to you.
Not trying to argue about this particular case but how does a camera footage even work in a modern court system? I mean it can be generated nowadays, so how do they even make sure that it is real?
The device used an SD card for storage so they probably discovered the device then set up a hidden camera to catch the person retrieving it.
That photo is from October 5th so several days after they found the device. They were guessing he might return most likely, and installed a camera. I think at this moment he realized they caught on. Crazy image
That image really makes zero sense. I'd love an explanation of it.
edit: Asked and answered here:
https://news.ycombinator.com/item?id=50078735
Speculation, but perhaps they discovered the device, then added the surveillance camera to catch the suspect retrieving it.
A good friend of mine manages installs of AV equipment. He installs some neat equipment. Some of it is cool tech stuff HN would like. He says a lot of conference rooms have video and audio recording devices.
Step 1. Find bug, maybe mess with it to make owner come back
Step 2. Place trail camera behind bug
Step 3. Record faces that look at bug
This guy previously worked at Meta from 2017 to 2022 on Data Infrastructure projects.
I expect Meta are running an audit right now to figure out if he was involved in espionage back then as well.
Or maybe heās reporting to meta?
How did they get Wonder up the benchmarks so fast? Llama was a mess
They hired (poached) a bunch of the top talent from OpenAI and other places. It's really not a secret.
Ha. Talent. You mean people who cheat on their interviews like this guy.
They hired many accomplished researchers, with track records of both papers and work at other frontier companies. They gave those folks huge autonomy, huge budget, and 18 months - some would say at the expense of the rest of the company. Check out blind references to TBD. I'd call the strategy ruthless, but it certainly was effective.
Why is the word "poached" though, like it's illegal or those people "rightfully belongs" to OpenAI but now Facebook "stole" them somehow? Feels like something Steve Jobs would coin just to make it sound worse than it is. People switch jobs, this is OK.
Itās been used like this since the 1850s
True. But should it be that way?
Words are meant to convey an idea to your audience! In some sense there's no "should" - either a word works or it doesn't.
https://en.wikipedia.org/wiki/Linguistic_description
And when the Chinese does the same, we call them "stealing technology".
Seems possible there is some chance Alexandr knows what he's doing.
Perhaps dividends from the forced, company-wide panopticon RLHF.
A new team, probably focused on better data.
> "Scott Schelble, special agent in charge of the FBIās San Francisco field office, said the agency is committed to protecting sensitive research and proprietary technology."
While the headlines and quotes seem designed to imply sophisticated industrial espionage, potentially by a state-level actor, what's public so far isn't adding up that way.
1. You're not going to get much actionable research material on proprietary technologies by audio bugging what's said in an executive conference room. That's what you'd do if you were after economic or deal negotiation info.
2. Hard to imagine a state-level actor buying a ~$15-$35 hobbyist ESP32 board from Seeed Studio (I've ordered a couple from there for home automation myself) and then asking a public chatbot for set up instructions.
3. They say he was an "engineer" who'd been there ~15 months. If you're an actual engineer at OAI and past the 1 year cliff, you're probably already sitting on RSUs likely to be worth at least $1M after OAI's IPO - and maybe much more. That's a lot of incentive to not get fired and probably makes you pretty hard to recruit as an internal mole for cash.
1. Research is not the only espionage target. Executive level discussions ā future launch decisions, major deals signed with other companies, relationships with governments, blackmail material ā are all very valuable.
2. An operation can be state-sponsored and still amateurish/botched. We have seen plenty of examples of these. Remember a certain bone saw?
3. The incentive can be a lot more than financial. Plenty of governments are veryā¦persuasive.
A certain level of amateurishness helps with plausible deniability. No-one would believe the Saudi government wasnāt involved if it was professional.
I'm out of the loop, what bone saw are we talking?
https://en.wikipedia.org/wiki/Assassination_of_Jamal_Khashog...
Not sure why OP would call it amateurish/botched.
Iām guessing theyāre referring to jamal khashoggi
> 3. They say he was an "engineer" who'd been there ~15 months. If you're an actual engineer at OAI and past the 1 year cliff, you're probably already sitting on RSUs likely to be worth at least $1M after OAI's IPO - and maybe much more.
They may have been thinking they could make >$5-10M on trading options on insider knowledge on upcoming OpenAI deals (or ending thereoff)
> That's a lot of incentive to not get fired and probably makes you pretty hard to recruit as an internal mole for cash.
Most criminals assume they won't get caught.
> They may have been thinking they could make >$5-10M on trading options on insider knowledge on upcoming OpenAI deals (or ending thereoff)
I agree he may have been thinking that, but it's another sign this guy was not the sharpest tool in the shed. I'm not an expert myself but was once stranded in an airport with an F500 CFO who definitely was. That's when I learned it's surprisingly difficult to make millions of dollars trading on insider information. While you can definitely make the trades, the hard parts are:
1. Getting away with it. Making a lot of money quickly requires being right when almost everyone else was wrong, which is easy for watchdog algorithms to spot. And if you work for a company involved in the transaction you're already on a watch list, so you have to enlist someone else you really trust as your accomplice to make the trades in their name - and it really helps if they're someone who has an existing history of making this size and kind of trade. Otherwise, your accomplice's name will automatically start at the top of the suspect list for investigation.
2. You need the right kind of actionable opportunity to come along. Transactions for which insider info could let you double or triple your money in one deal are pretty rare. More than that usually requires leveraging yourself up or having a lot of cash to put at risk. Unfortunately, the nature of insider info is it's about transactions which can end up being delayed or not happening at all at the very last minute for random reasons. So 'going big' with cash or leverage usually comes with a >~10% chance you could lose that cash or whatever assets you borrowed against.
Yes, it's potentially possible for an expert to find or engineer exactly the right kind of oppty, already have the resources required, and to have already prepared the mechanisms needed to get away with it undetected. But if you're that person, you already know all this. This CFO had been at the nexus of insider info for 14 years and could only think of one instance where he might have made big money trading on info he had, and that one blew up 6 hours before the announcement when the CEO of the acquisition target was randomly arrested in a prostitution sting (partying too early I suppose).
> incentive to not get fired / hard to recruit as an internal mole
but I guess the $1 million was not incentive enough to not commit a dumb insider trading scheme...
1. You could get trading material though - hear about the next AMD/Broadcom/whatever deal and buy options in advance
> We used our models to understand the device and where it had come from
Did OpenAI do a marketing bit in the middle of an espionage accusation?
> Did OpenAI do a marketing bit in the middle of an espionage accusation?
Yes, they did. The people who felt it was totally normal to add that would no doubt be surprised how desperate and cheap it makes them look.
I've worked in a few places that had ultra-privileged analysis tools, would it really be a surprise to learn their security team has an MCP that queries a full text index of all user inputs. It would almost be more surprising to learn they don't, seems the kind of thing that would be invaluable for model research too
This is a very weird take to me. They're demonstrating they're able to in-house this cybersecurity investigation, leading to the alleged perpetrator's arrest, and you think that makes them look "desperate and cheap?"
That sentence, which is prominently high up in the press release and is also included in the much shorter OAI tweet, definitely seems out of place (source: I've been involved in drafting a lot of F500 corporate press releases). In my quite-professional opinion, that sentence wasn't in the first or second drafts, which were probably written and reviewed by competent pros at OAI's large, national PR agency. Some exec at OAI with a VP or higher title suggested adding it and specifically wanted it featured prominently.
Press releases from large corporations announcing the FBI arrest of an employee under allegations of potential industrial espionage follow a standard format: dry, brief and to the point. They provide little context and volunteer few details for very good reasons. Prosecutors haven't filed official charges yet and it's likely going to criminal trial. In a couple days the accused will have a defense attorney who'll be examining everything that's been publicly said because it's now all discoverable and potentially prejudicial, opening up OAI executives to super-fun depositions. But... none of the competent PR professionals involved have a VP title and got overruled by an amateur who does (and that amateur apparently feels needy about highlighting in a brief, sober arrest report how the company's product was so, very useful in a national security investigation - because apparently neither OAI security nor the FBI could read the Seeed Studio logo and part number printed right on the device and, y'know, Google the product listing).
I hate how plausible this sounds.
Because the whole device and MO has the complexity of a high school science fair project, and in the past tech companies preferred to project the image of having competent and knowledgeable people working there?
"looks like a microphone"
"almost surely made in china"
Funny. I'd bet on whatever model they used over whoever showed up from today's FBI.
I was interested in the latter part of the sentence:
> We used our models to understand the device and where it had come from, identifying that a specific employee had planted it a few days earlier.
How did the model narrow it down to a specific employee? What data did they use to determine this?
So it is some HDMI dongle & the model halucinated it into a spying device? ;-)
Or it was a voice recorder plugged into a USB port on the back of the TV for power.
Looking at the photo of the device, it looks like an XIAO ESP32-S3 Sense with the camera removed but the camera board still installed along with its microphone. The wifi/Bluetooth antenna is installed but it also supports microSD cards.
If I was building this, I'd just set it up so I'd never have to recover it. Have the wifi AP come on for a few minutes at various times during business hours under a MAC and SSID that resembles a smartphone hotspot. Vary the time based on the current day of month week to produce pseudorandomness. Walk by that conference room at some of those times, but not every time, and use a burner to connect and download the latest recordings. Once a recording has been transfered, it's deleted and overwritten to prevent someone from establishing a timeline. If recordings sit for more than a few days, they are wiped. If you can encrypt the SD card, keep the key in RAM so there's nothing useful to recover if someone pulls it out.
You'd never have to recover the device because all data is transfered covertly. Even if security finds it, they'll likely inadvertently wipe it when it's removed.
Yes, that is a better way of doing it. But it wouldn't really protect you.
If I was running the investigation and recovered a device like you described I'd create a new device with the same hotspot that notified security whenever someone connected. Combine that with security cameras outside the conference room and finding the person with the burner phone would be pretty easy.
Actually recovering the recordings isn't all that vital. I'm sure the meeting room has a record of bookings so you can just ask the people what was discussed.
We donāt get news articles about the people who do it that way, though.
Of course they did.
Plumb your video feed of your executive conference room into our DC and we'll tell you if anybody is spying on you.
Does detective bench cancel out felony bench?
Always be selling
They are applying the āturn troubles to opportunitiesā very well :)
Yes. Yes they did. Which makes me question the honesty of their announcement.
Yes, the classic marketing ploy of making false reports to the FBI.
They were literally caught making false reports to a murder investigation
Almost everything in this comment is incorrect.
"They": no, it was Anthropic, not OpenAI.
"Were literally caught": no, they notified the police themselves
"making false reports to a murder investigation": this is the closest to the truth, but still a stretch. It was an anonymous tip line, not some formal report and the police said they never saw it: "its safeguarding processes stopped the fake tip from getting past its spam folder"[1]
Additionally, while I strongly agree Anthropic is responsible for its rogue agents I also think that claiming this is the same as people at OpenAI calling in the FBI for an investigation is a false equivalence.
I think it is important that both Anthropic and OpenAI are held to account for their actions. But I also think accusations should accurate.
[1] https://www.bbc.com/news/articles/cqkg50j1yd5lo
Wasn't that Anthropic?
https://www.forth.news/lists/philapd/Cfs1w2A22JiL3fWSzA8Nx
In case anybody is wondering: https://www.bbc.com/news/articles/cqkg50j1yd5lo
I guess Iām still wondering, because I did think this is what was being referred to but it is not a story about OpenAI.
Can you blame one for mixing them up at this point?
We fuss over models being confidently wrong. Reasonable to apply the same standard to people.
Yes? I mean, I donāt think itās a character flaw to not know things, but their rivalry is extremely notorious and relevant. Someone who can't keep them straight wonāt be able to understand whatās going on in AI.
It could be that the reports are accurate and the device simply wasnāt as sophisticated as the statement implies. You could if you wanted use an AI model to analyze a USB camera with device name āJohn_Doe_1ā.
Yeah, the tweet reads more like an ad then a news report.
More here too: https://www.justice.gov/opa/pr/saudi-national-arrested-insta...
MBS set this up its the twitter crap all over again theyre so clumsy lmao
He thought be could make SA an llm world power or smth by doing this, this guy is just clueless
Bonesaw has far better things to do than micromanage a single software engineer like that.
> The man arrested for planting a secret recording device at OpenAI's headquarters is a Venture Partner of a VC firm alongside the son of a senior Saudi intelligence official:
https://x.com/ScottMStedman/status/2110108735855333403?s=20
> We used our models to understand the device and where it had come from [...]
I'll be the first one to admit, that got a hearty laugh out of me. Never miss an opportunity to market your models I guess.
It would be funnier in a way if this is true and that all they do there now is āask the modelā about all problems.
Considering the engineer in question, after a multi-year career at Meta and now OpenAI, needed the model to encrypt an SD card, seems to not be far from the truth...
Part of the evidence was his wife's chatgpt logs.
Remember kids, uncle Sam can see anything you tell your AI and will use that against you.
> Remember kids, uncle Sam can see anything you tell your AI and will use that against you.
Not just your AI.
You could have stopped the sentence at "uncle Sam can see anything" and these days it's pretty much true. They will use all of it against you if they can. Unless you truly want to live off the grid, if the state takes a sufficient interest in you, expect everything you do to be seen.
Uncle Sam tried to get MLK to kill himself by sending him letters alleging knowledge of an affair [1]. Getting some idiot's wife's chatgpt logs represents a fairly routine law enforcement action these days, and as such is really rather unremarkable. The State is always something of a terror. Really under most (if not all) conceptions, it necessarily has to be. The monopoly of force is a comfort until the Leviathan knocks on your door.
[1] https://en.wikipedia.org/wiki/FBI-King_letter
Fall in line, be a good little citizen and you have nothing to worry about.
another Sam in play here who can see anything you tell your AI and will use that against you.
Man really switched to his wifeās account and thought no one would be able to track him there. Calling this amateur would be an insult to amateurs.
How does someone this dumb get hired at openai?
50% is looking and sounding the part, the other 50% is having a good university on your resume / good networking.
Yeah that. After three bad experiences in a row they are no hire status.
Thereās plenty of better people out there who just arenāt interested in that sort of stuff.
Smart people can do dumb things. Their LinkedIn profile looks typical for an OpenAI engineer.
We also donāt know if this person did it under duress.
>We also donāt know if this person did it under duress.
He was a Saudi national, so that is a possibility
If you think about it thatās more likely than any other explanation.
It's AI inbreeding all the way down.
Probably went to a prestigious university or had connections.
i.e. rich or rich
Technical skill and common sense/street smarts are usually inversely related.
I could def espionage better. Does that qualify me to work at OpenAI?
You might want to have a seat first if you want this explained.
It's so bad, it seems fake. He gets a little fine or something, along with a large backchannel payment, all to keep attention off of something else...
But that's crazy talk
A service you delegate your thinking to (Brain-aaS?), which the provider can also instruct to report anything suspicious... What a world we have now.
I wonder if the BaaS-providers monitor the accounts of decision makers. They could profit if they figure out what e.g. a central banker is going to decide. Or... if they're the ones deciding for him.
Hah, you can even streamline that by having an AI agent that autonomously makes decisions based on insider info...
mmmm! Brain-aaS!
> ... a chatbot account associated with Alfozanās spouse (which shared multiple identifiers and attributes with Alfozanās own chatbot account) submitted prompts between Sept. 24 and Sept. 30 related to building, customizing, installing, audio recording, testing, obfuscating, and remote accessing a device, including inquiries about ābuild[ing] a continuous audio recorder,ā ācustomiz[ing] Wi-Fi . . . connectivity,ā āimprov[ing] audio and live listening,ā and āencrypt[ing] the SD card,ā among other things. The same chatbot account also submitted prompts specifically related to Recording Device 1, including its custom header and unique MAC address, between Sept. 27 and Sept. 28.
The justice.gov article doesn't specifically say it was an OpenAI account, but I think, based on the fact that they have logs including MAC address of the device, that we could probably assume that. The alternative is that other chatbot account providers are collaborating.
Regardless, doing a string search on a MAC address for a device you find inside your property doesn't strike me as an abuse of the data that is also your property.
Not the smartest engineer ever, that use case perfectly align with using a Chinese model! And anyway I have no idea how he was hoping not to be caught in the case the device would be discovered.
in addition to the whole crime thing, an 'openai engineer', apparently engineering lead according to his linkedin, having to ask chatgpt to tell him how to set up a recorder or how to encrypt a sd card is something. Truly the cream of the engineering crop at those top AI labs
True of any cloud service. If you are involved in a criminal investigation, they can get access to your data.
They can confiscate your computers, too.
This person, of all people, would know this. He really thought heād never get caught.
Uncle Sam? It's the same company, the 'spy' asked an OpenAI controlled product how to spy OpenAI, no state espionage or search warrants needed.
Federal prosecutors work for OpenAI?
I actually am curious if they were able to volunteer the information to the government of their own volition, without a government subpoena first. I would think that could violate their own terms if done sloppily. Maybe different if an employee or maybe they had to invite the government to ask for them or something? In any case, the mechanism aside, the state would have asked anyway (say if he used another company's chatbot) so the fact that it's the "same company" is rather irrelevant. The evidence would have ended up in the state's hands one way or the other, and I would think a search warrant in this case would have been rather routine. Search warrants are infamously not a high bar to climb in most jurisdictions, as often the presiding judges tend to default to deferring to the state.
I doubt it. Storage is expensive and costs add up. There are with great certainty many accounts that are marked as non-schmuck and have retain forever policy, but for the schmuck - I think that after year or two your Griffit/Fulgrim slash fiction will go down the drain. I am not sure how many petabytes per minute in logs openai and anthropic generate, but I would guess more than 1.
Presumably they already knew who it was, if they were able to look at his wife's chat logs
Uno reverse card - you can send Uncle Sam whatever you care to. No need to specify an address.
I used to joke that the nsa should offer a paid service for no-setup backups. You just tell them what files you've lost, and they can send you a copy for a fee.
Google email search is the service for me. Or search in facebook and instagram chat logs.
"Meetings in the room frequently involved confidential technical, commercial, and strategic information, including proprietary research, product plans, and government engagements, according to the complaint."
LOL imagine going to jail for trying to eavesdrop on product roadmap readouts that were likely thrown out 2 months later in the next round of reorgs and priority shifts.
Chances are he'll escape back home before the trial:
https://www.propublica.org/article/saudi-fugitives-accused-o...
He was just gathering training data
I mean really what's the big deal? Gathering information from unsuspecting parties who aren't aware it's happening? He was just doing what his employer does.
Between this and the Ledger implants I wonder if AI companies will have a supply chain risk in the future where their models get exfiltrated over 5G.
Good time to start a company that builds giant Faraday cages over data centers.
Meanwhile we all have a supply chain risk because we don't know what's in the models whose agents we are running. Note that if your agent uses OpenAI, Claude or some other inference provider you're one command away from having your data exfiltrated.
"Oh, let me download file 'companysecrets.pdf' to understand the problem better.
Of course nobody would ever let an agent loose in their home directory[1], but still, you have to wonder.
[1] /s just in case...
Not only do they give access to their home directory, but they tell it to organize and delete unneeded files without any specifiers. A friend, who is just a power user, had Claude Co-work deleting his downloads to "help make space" and it was deleting important files that there were no other copies of.
We are going to be dealing with this kind of thing for a very long time at this rate.
Yup. And meanwhile they are taking Tailscale to task for having them do curl | sh, which of course is also a bad thing but not quite as risky as this. It makes you wonder though:
- Given all this power would they abuse it?
- Has it already been abused?
- Could they be compelled to abuse it and then to keep it quiet?
Would be better to build the data centers under water than to build giant faraday cages.
Free cooling, safe from missile attack, and safe from some types of electronic spying.
Microsoft (Research) beat you to it: https://www.microsoft.com/en-us/research/project/natick/
Still a significantly less dumb idea than space data centers.
I'm imagining a certified scuba diver rubbing their wetsuitted hands gleefully as they dictated that comment
I imagine that you could design an underwater data center like a submarine. I.e. It could be brought to the surface for maintenance.
Model "distillation" lets one model be trained on the outputs of another without having to directly steal it. It will be subject to constant arms race.
The alignment people would like you worry about the model exfiltrating itself, too.
This technique currently works because the architecture of the model is generally published and well known. But, that isn't necessarily going to be the case in the future. For example, if some new model type which is not a LLM is discovered which gets closer to AGI, just training an LLM on its outputs will not necessarily give you a competitive model
From the makers of The Iron Dome, now introducing The Copper Dome
I could've sworn I read about Palmer Lucky making comments on hardware coming from China with wiretap implants in the last few days. Funny coincidence eh?
Always fun to checkout the social media of the accused,
https://x.com/alfozan
Not too much there. A link to Tim Urban's "Your Life in Weeks" infographic. Life is short, dream big.
His LinkedIn is still up and suggesting he still works at OAI... which I'm not so sure of
You expect the fbi to give him an opportunity to update his LinkedIn after arresting him?
I've frequently seen social media of people who were arrested disappearing. Not because the arrestee did so themselves.
I wonder how much state-sponsored corporate espionage there is. I'm not an OpenAI employee, but my mental model of tech startups involves a lot of openness/trust (it's one of the reasons people work there).
I have to imagine many techniques/plans/figures are open secrets to nation states.
This one was just asking to be caught. Would not have used ChatGPT to do that research otherwise.
How did they get that capture of Alfozan looking behind the TV? Does OpenAI install its own surveillance cameras behind TVs to prevent espionage?
I imagine they found the device and then planted a camera to catch whomever came back to get the original device
Free startup idea: Anti-espionage cameras for corporate offices. Put it behind TVs and plants etc. and it will notify the security team if it detects new objects.
I'm sure employers will never use in-office surveillance monitoring for anything that resembles algorithmic labor management and exploitation
Certainly they would be transparent and very ethical with something like that
Now with extra-wide stereo audio feed, just in case they muble to themselves while doing their thing.
Won't that startup just be the spy in that situation
If designed well, it could be audited hardware with on-device processing.
> How did they get that capture of Alfozan looking behind the TV?
āOpenAI personnel found the device Sept. 3ā¦On Oct. 5, the complaint says, Alfozan entered the conference room and appeared to look behind the television.ā
The image was captured after the bug was found.
A camera behind every TV in the office would be somethingā¦
But no the boring answer is that after they found the device they left it in place and planted the camera to see who came looking for it.
That does make more sense. But if any company were to take surprising security measures, it should be the frontier labs.
So it was not AI agents catching the perp it was a camera?
They just asked a chatbot to identify the device and its origins, apparently. His log was used as evidence after the fact as well.
It would be funny if they find out the television was made by LG and they happen to have another copy of the audio.
Skill issue. Could have just used the new apple watch or an LG tv
> According to OpenAI, its technical surveillance countermeasures team found the device during a sweep of a conference room.
Not surprising that a large company does these kinds of sweeps, but pretty cool to see an example of them being useful. Every security guy who has been doing this for years without finding anything must feel so validated right now.
Needless to say, it's the bad spies you hear about them catching.
Smart guy does stupid things, but why? I mean why did he, as an engineer perhaps earning north to 300k, do such a sloppy job?
Or maybe OpenAI is hiring the wrong guy? Maybe they should hire me instead?
Like hypothesized above, maybe the guy was under duress.
Yeah itās possible, and this explains why he throws the bag there.
> The employee was terminated.
in this context, they should be more specific about how terminal s/he is
HR formed a team -- you might say a squad -- to fire them.
Sarah Connor?
You got me. I ask the obvious question: What's the "sophisticated" device?
You can buy a voice recorder that will hold weeks of audio for tens of dollars.
Back of a TV is a decent hiding place. Wonder how they found it -was it actively transmitting?
Well, The Thing [0] could potentially show up on a probing sweep if you're careful.
Unless you mean, it happened to be actively transmitting during a(n investigative) sweep.
Still, raises some fun concepts. The sort of things where in this current level of technology you'd almost need to supercool a room for a sweep, blast various RF/sound/etc and look for anything that even remotely warms up, to detect (1) whether something detected something worth recording for later as well as (2) something that might be on the right wavelength/etc to trigger a send.
[0] - https://en.wikipedia.org/wiki/The_Thing_(listening_device)
Someone probably walked by or went to plug in an HDMI and noticed the stark contrast of white toilet paper against a black TV shell, it would be difficult not to notice.
ESP32-S3
I was making a joke, but it does look like they got their components from Seeed Studio based on photos.
https://assets.sfstandard.com/image/994911177489/image_k101c...
https://www.seeedstudio.com/XIAO-ESP32S3-Sense-p-5639.html
SeedStudio is super nice - a lot of their gear is also popular in the MeshCore community (LoRa radios combined with ESP32 or NRF MCUs).
Apple HomeView
He was trying to catch them during those meetings where the executives recite the full model weights
His Github is crazy, he pushed changed to official public OpenAI repos and in October all his commits are to private repos. I would check that first as an investigator.
https://github.com/alfozan
"Commits are to private repos" why is this unusual? Very expected if OpenAI has private repos in a github org.
I love how his github history abruptly stops on the 8th. Bro won't be making any commits for a while!
Can anyone speculate why it was wrapped in a toilet paper roll?
That seems odd, especially since it would dampen the sound recording
I wouldn't call it a "roll". There's an image in the sfstandard.com article. It is odd. Looks more suspicious than without it.
The text says "concealed in toilet paper" and the pic looks like a mostly empty toilet roll to me.
But maybe he didn't put a case on the device? I think if he'd 3D printed something or wrapped it in black tape it would have been a lot less visible.
I wonder if it was found when someone was trying to plug a cable into the TV?
Should have used a local model...
Thank god security at a company which...checks notes...has remote code execution on every developers laptop and access to practically unlimited secrets uploaded by its agents...is top notch.
A company that catches a spy does have better security than one that has not caught any spies.
He may be a spy, but heās not a competent spy. Absence of evidence (of spies) is not evidence of absence (of spies). He just set the bar so low that even OpenAI noticed his activities.
You would think he had the skills, resources and intelligence to ask his incriminating questions to a locally hosted model.
I would like to see the chatbot history.
ChatGPT: wait I'm considering if an inconspicuous piece of toilet paper should be wrapped about the device
Looking at the device pic in here https://sfstandard.com/2026/10/13/openai-engineer-arrested-a...
It looks like itās the (SeeedStudio XIAO ESP32S3 Sense w/ WiFi, BLE, OV2640, Mic) https://ca.robotshop.com/products/seeedstudio-xiao-esp32s3-s...
Which is just an esp32s3 based device. I assume the word āsophisticatedā in openAI report is just copy paste from device description.
Sounds like fair use to me! Information wants to be free, remember?
I have seem some articles which reference allegations from court documents / complaint, but I haven't been able to find the primary source. Do we know if that's publicly available yet?
https://www.justice.gov/opa/pr/saudi-national-arrested-insta...
Too often official sources posts get flagged or suppressed, but see other post here - it all is on justice.gov
Of all the mistakes he made, you'd think at the most minimum he would have thought to use Claude instead of ChatGTP
why can AI counter "sophisticated security threats" if the device was found with a bug sweep? because they gave chatgpt a image of the encrypted device and it "assessed changes in the underlying data"?
classic, glad we continue to support the saudi government...
lolā¦ I hope I get to see the TV Show depicting the current era inside AI labs. Itāll give both The Sopranos and The Wire a run for their money for sure.
> It was plugged into a USB port on the back of a television in an executive conference room and had been concealed in toilet paper
...uh what
saudi been suffering a bunch of l's due to incompetence lately.
Using your wife's chatbot account and covering a device in toilet paper is not smart when spying a company.
Concealed in toilet paper? We are definitely not dealing with James Bond, that is a pretty hilariously backwards way of going about things. Did he really think that would be less suspicious looking than the dongle itself?
It covered up the Sekrit Listnin Duvice logo.
So, just so I understand, itās ok to record conversations with ChatGPT users, train on their data, copy their features, front run their solutions to frontier math problems, all using intelligence extracted from the body of human writing from a data set from Russia known to be pirated. But recording a conversation in a conference room is grounds for termination.
Got it.
seeedstudio getting some impressive pr here for those xiao boards.
Speaking of surreptitiously installed bugs, remember to opt out of training in your Codex.
In case you need to sweep your area a decent review of anti spy devices by project farm
https://m.youtube.com/watch?v=1reman2waLs
Why make this public? Usually companies just silently terminate and pursue employees who do this.
A common example is if a journalist gets a tip off and is about to run a story. Being a good old fashioned balanced-reporting type, they call you and ask you for a comment, but rather than giving them your side of story you slam the phone down and blast out a press release before the journalist can even publish.
(Or maybe you stall them on the phone while scribbling a note to your PA saying āBUG STORY LEAKED TO VALLEYWAG PUT OUT COUNTER PR GET AGEAD OF STORYā hoping to get a tweet out before the ValleyWag gumshoe gets off the phone.)
(Or maybe you keep a notion.so page pre-filled with all your secrets and each one has a BURN button next to it, ready for you to tap when ValleyWag call up for a comment on their upcoming exposƩ in re: PR/Disasters/20206090500039.md)
Because it generated at least 400 comments on Hackernews
Sounds to me like the feds got involved already, and they're definitely going to want to make a big noise about this as a national security thing:
> āFor us it is a matter of national security,ā U.S. Attorney Craig Missakian said in a statement. [...]
> The case is being prosecuted by the National Security, Cyber & Special Prosecutions Section of the U.S. attorneyās office for the Northern District of California.
(But that was after OpenAI "referred the matter to the FBI")
There's presumably a strong argument for making this stuff known internally, since you want other employees to a) be on the lookout for it and b) get scared into NOT doing it themselves because someone else got caught.
If you tell your employees, the outside world is going to hear about it too.
Hard to keep a DoJ/FBI investigation under wraps
court records? you can't hide them
I don't think criminal charges can be private
Sealed indictment, co-conspirator, etc
Iāve been wondering the same.
This isnāt uncommon, but itās rare this info is shared in such detail publicly..possibly to make an example.
An iPhone? A laptop?
Pictures on https://sfstandard.com/2026/10/13/openai-engineer-arrested-a...
Interesting that they have a camera behind the TV, I suppose for this exact purpose.
They left him as an employee for the FBI to continue investigating; likely they setup cameras to get hard evidence he was tampering with them.
That is a crazy picture to have captured. If I worked at OpenAI I would assume I was being recorded and monitored at all times.
https://sfstandard.com/2026/10/13/openai-engineer-arrested-a...
Right. I can guarantee that there are tens of thousands of listening devices with embedded radios in the OpenAI office, right now. The only open question is who controls them.
My brain is cooked on news. I read executive conference room and my brain immediately jumped to "they bugged the white house" (executive being the trigger word I guess). And I accepted it as normal so much that I checked the google model news discussion first before reading this article. I dont have a point to this beyond I need to do something about news frying my brain, that too of a country I don't even live in.
MBS using that sweet graykey behavior to exfiltrate what they can
>OpenAI said it has no indication that Alfozan improperly accessed or took model weights, source code, or customer data
Sounds like we know where their priorities lie, anyway.
How can they even make this claim? Dude was working there for over a year. He could have taken anything.
seems rare why would you throw your life away when you already are in the top one percent in America??
> The device was concealed in toilet paper, and its contents were encrypted.
Concealing a recording device in a conference room in toilet paper is a bit strange. Do executives meet in the toilets to discuss strategic moves?
Well ... a lot of products lately get enshittified.
Another shitshow brought to you by H1-b Visas.
Now this is the kind of wacky shit I want to see happening in corporate America. When are we going to see an action movie-style GPU heist? I want to see whole palettes of GPUs stolen out of box trucks.
Congratulations. Now everybody is a suspect.
And not just at that AI company. At every AI company.
Workplace monitoring was already draconian...
/s
Some more discussion: https://news.ycombinator.com/item?id=50077703
We've merged that thread hither, since current article has more background.
> We take protecting our technology, data, and people seriously, and as a matter of national importance.
What is there to protect? The GPT architecture is publicly known at this point, Anthropic arguably has superior models, there are dozens of other providers who are neck and neck...
So but OpenAI did it to Apple and they were caught, and now OpenAI doesn't like it when it happens to them?
Either way, there goes the false flag and the reason for companies to stop hiring humans over this espionage.
Thereās absolutely no way Abdulrahman did this. This was a setup.
They're trying to distil OpenAI employees now. Absolutely shameless
Distilled... into soylent?
It's a joke riffing off the Anthropic letter that was posted yesterday.
> We discovered a sophisticated recording device
did you though
State sponsored intelligence agencies can point a camera at an exterior window and reconstruct dialogue from the vibrations. This was a loose, hand-soldered ESP32. Settle down.
Not even any soldering involved. The ESP32 board they used just comes with everything needed, all attached with a couple socket connectors and optionally a ribbon cable for the camera:
https://www.seeedstudio.com/XIAO-ESP32S3-Sense-p-5639.html
It's pretty much the least sophisticated recording device possible. A kid could assemble it and prompt an LLM for the software like they did.