Wow I missed this story entirely. I'd seriously considered moving to OpenRouter at my next external AI renewal due to the breadth of options and pricing but this has absolutely ended that consideration.
It's clear from the breadth and variety of exposures that there's no due diligence or guardrails inside OpenRouter on who they market and how they classify their providers. Their response to this indicates a few first steps but none seem to have the force of law and i'd have expected more from a Stripe aquisition.
Very glad that my daily driver is a local model and I only use these external services sparingly.
That being said, the best ZDR agreement is the one that your company has directly with the inference provider(s). Unfortunately:
- Not everyone is large enough (or has enough $$) for that.
- Ultimately youâre in the same position of having to âtrustâ that theyâre doing ZDR; even after auditing their security docs.
Yes that's why Revolut (Europe's largest digital bank) has just started self hosting open weight models using 56 Nvidia B200 GPUs. So their 2000 engineers will use that for most things.
Thatâs not true. They switched for cost and productivity reasons.
They explicitly specified USA, EU, UK, and Israel-hosted services as acceptable as long as itâs ZDR. They said that inference hosted in China was something they didnât want.
I generally trust (for example) bedrock to honor the ZDR agreement as written and to have secure infrastructure. However I don't fully trust the routing service I use to access bedrock to have rigorous security practices.
> Ultimately youâre in the same position of having to âtrustâ that theyâre doing ZDR
Could it be a moat to offer a zero-trust (based on technical impossibility) ZDR solution for corporate customers, or is everyone whoâs paranoid enough about that just going to host local models on-prem eventually?
Not an expert, I consult with medium size firms that so far have seen nice things from the AI. However, the lightly-informed C-Suite hears about the hacks and the data threats on CNBC/Bloomberg which has made them cautious. Even with their cursory understanding they are saying "can we look at these open models". N=2
Isn't that the problem themselves, in that those issues only exist if you go for open models, but outsource inference to third parties? It seems to me that there are only two safe options: either go full local and do inference on-prem/in your own cloud, or just deal with the big labs directly. Everything between these two extremes involves third parties with sketchy reputations.
Unfortunately, I highly disagree. I'm deeply disappointed that they don't even have the gall to call the violators out so we can be informed. Dug around for so long and I can't still find all the rule-breaking providers. It's a bunch of "trust us bro" nonsense. Name and shame should be the minimum here.
If they are going to name and shame they have to put in a lot more work first to be sure their evidence is tight enough to dissuade or dismiss defamation lawsuits, and there would still be risk and lots of drama for people too busy for that shit. It just isn't realistic to expect that in my opinion.
Its possible to be very certain of the evidence you have but not be able to prove its not been tampered with or fabricated, for example.
what move? OpenRouter is garbage, how could Stripe have thought it was a good idea to buy them? $7b for revenues that will evaporate, inevitably, all at once, in weeks!
I was affected with just one prompt by this incident, which is non issue to me because I was trying different models at that time so there was nothing serious. The troubling problem I see in this incident is the âprovidersâ in OR just casually forwarding prompts to upstream (MiniMax) instead of providing inference service them selves.
Coincidentally later on I set my OR settings to only allow some providers that I can rely on, not blindly let OR route to wherever they like.
For those who don't know this is the first thing you have to do when using openrouter, change guardrails settings from blacklist to whitelist, pick 1-2 providers you trust (do your homework, or let $agent do it for you), and if you want to go the extra mile pick the fastest/cheapest provider for you model (again do the research). The defaults can't be trusted if you care about privacy.
I got the email. The tl;dr is MiniMax (the Chinese AI lab) is actively reading your prompts, and a security researcher discovered this by putting decoy URLs into their requests which were accessed by MiniMax.
More importantly, many providers (AtlasCloud, Novita and StreamLake were named but there seems to be more implied in the blog post) have been rerouting requests to MiniMax servers on the backend and not processing their own prompts.
The obvious question is whether OpenRouter has done any auditing on their inference providers. Based on the researcher's blog post, I guess not.
> A routing configuration issue caused an API forwarding path intended for specific customers to apply more broadly.
> If you used MiniMax models through Novita between 2026-05-28 and 2026-11-11, your request contents may have been accessed or retained by a third party. This includes requests made directly through Novita and those routed to Novita through OpenRouter.
ârouting configuration issueâ?!? Imagine if AWS got caught storing your S3 on some other providerâs disks and they said it was a ârouting configuration errorâ! They shouldnât have software that can do this at all! The API is supposed to be for their own GPUs!
Togetherâs is weird:
> As part of routine software testing, we collected prompt samples authorized for that use and stored them in a private Hugging Face repository. The samples came from requests to the MiniMax M3, GLM-5.2, and Kimi K2.6 models, primarily on June 24, July 4, July 8, and July 14, 2026.
Did Dragos really âauthorizeâ Together to store their canary âprompt samplesâ? Somebody screwed up here.
Together.ai has made a public announcement that they are fucking clowns who cannot be trusted with anyone's data.
> During development of our public GitHub repository, xorl, an AI coding agent committed a credential that provided access to the private Hugging Face repository. Our automated scanning detected the exposed credential, and we reverted the commit.
They didn't rotate the credential they knew was exposed!?! I guess "an AI coding agent" didn't think of doing that. Also FYI kids, reverting a commit does not delete the data from the repository.
In before they claim an AI agent wrote this disclosure that included certain "errors".
This is real classy coming from a neocloud with a $8.3B valuation. Shouldnât a neocloud with a bunch of Nvidia GPUs and a bunch of investment (including from Nvidia) be using those Nvidia GPUs to run inference in house? Isnât that the entire point?
Amusingly, the first time I typed âneocloudâ it got autocorrected to âbro cloudâ. I almost decided to leave it alone.
This and the way Together seems to be framing it in their disclosure post ( https://www.together.ai/blog/security-event-prompt-leakage-f...
) raises a lot of serious questions about the integrity of Together in more ways than one.
Is the Together post really a fair disclosure that accounts for what Dragos identified? Their explanation is far from clear. It makes me wonder what the legal agreement was, and if the work mentioned in Together's "Polymorphic Inference" blog post, mentioning speculative decoding (using cheap models to draft) is a related practice they have routinely. Pretty worrying.
I'm sure some will make this into simply blaming Chinese companies for this (not you), which wouldn't fairly account for the risks of the industries optimisation incentives. (Together's HQ is in SF AFAIK).
You left out the best part from Togetherâs
âDuring development of our public GitHub repository, xorl, an AI coding agent committed a credential that provided access to the private Hugging Face repositoryâ
As an aside, I always wondered about decision models. They are so cheap. I canât help but think that providers must be using them as honeypots for data acquisition:
You do dont find real world company documents on the web that you could train on. And companies certainly wonât upload them to you so you can improve your models, not even with incentives. But provide a model with dirt cheap input pricing and free outputs - and you gut yourself a firehose.
To think that the âlabsâ wouldnât train on this is like thinking theyâd never have trained on shadow library books because, âwell, itâs illegalâ. They did it back then and it gave them an advantage over the competition. So why wouldnât they now?
> As an aside, I always wondered about decision models. They are so cheap. I canât help but think that providers must be using them as honeypots for data acquisition:
This one is actually the simplest for me, you can run decision models on your own laptop, and you can see how cheap and fast they are, it's actually amazing how "good" (jev etc is still better) they are.
So instead of thinking of them as a "LLM" for hosting pricing, it's easier to think of them as a API as a service, it's practically "how much can we charge someone for processing in bulk for an action and to get them sticky".
You setup whatever workflow you want for the decision models and then you practically never change it, because why would you, it's dirt cheap (for you and they provider), they (the provider) is making money on your outsourcing for every API action and it's going to be consistent because it's going to be for a long time.
But nobody knows how jev is actually built , what open source decision models? All the ones available kind of jerryrig it on top of a LLM last I looked into it.
IIRC JEV is known to be built on Qwen, you simply ask it questions to that effect. No idea where I saw the thread, but someone did this exhaustively using many variations and the answer was conclusive
Decision models can be cheap because itâs a fundamentally cheaper technology than LLMs. So that part isnât suspicious at all to me.
That said, if you donât believe OpenAI when they say they donât train on your decision API inputs you shouldnât trust them regarding LLM inputs either.
You should assume that no inference provider is honoring any public commitment they have made about data retention. You should assume that 100% of what you share with an LLM over the wire is being recorded, saved, and used as training data. Any other stance is indefensible naĂŻvetĂŠ for anyone who has been around the tech industry for more than, letâs say, 5 years.
There are real exceptions to this rule, like Tinfoil.sh who cryptographically ensure they cannot retain the data. (As long as one trusts the chipmaker, I suppose.)
Of course, even with a privacy-respecting inference provider, if a modern, relentlessly proactive LLM runs in a harness with internet access, it can easily accidentally exfiltrate your data anyway.
Rather you have to trust the equipment operator (which in this era of clouds isn't necessarily the vendor) not to intentionally compromise their own hardware (see https://tee.fail/). But other than that quibble yeah, tinfoil looks like an excellent option that's much less likely to leak than current alternatives.
read anthropics tos. it says very specifically that they do not train on "inputs" and "outputs" of customer data. everything in-between is openly declared to be their legal right to train on. in terms of their ability to extract intelligence from inference there is zero difference between this and training on "inputs" and "outputs."
By the definition of how these companies see data, a simple transform to restructure the input and output would be sufficiently different to pass this bar.
I fully expect them to be storing it, just the period of time they are storing it for is up for debate. Even everything older than six months ago, at this stage for me would be pure junk. If you want to train on that when I was still using soul files and barely any harness, heh, well that's your problem now!
They are not all lies, but these are definitely circumstances in which case lies should be expected because it is possible to profit significantly by making them. Lots of new companies springing up over night making promises they won't steal the valuable data passing through their hands.
The number of providers they have added is absurd, there have to be some that have almost no compute and just upstream everything. So stealing is their only possible business model.
openrouter.ai hasn't really promised that they do any vetting or due diligence, as far as I know. This whole space has had the scammy crypto vibe from the beginning.
ZDR through AWS Bedrock is still safe right? AFAIK Bedrock self hosts all models they provide. Even proprietary weight models like Claude are run within AWS network.
Is there any information on the three companies that violated ZDR? I could only find mention of one of those affected providers in the original article (that OR links in their post): Lark. Trying to figure out if I ended up routing to one of the affected "ZDR" endpoints.
I don't understand why these names are being coyly tiptoed around as if they're trying to protect the privacy of these companies blatantly lying about ZDR. Why isn't OR calling out the actual providers they off-boarded by name? Why is OR basically protecting them under a cover of anonymity?
Lost a lot of trust in OR that's for sure, in-spite of some other positive reception I see here for this move.
They state they notified customers who used any of the problem models via email (I got an email about it), which answers your question of whether you ended up routing to one of those endpoints. They apparently know who was routed to all those providers, so you don't have to guess.
That still leaves the desire of people to make informed decisions about what providers to use unresolved. If reputation is the only thing we have to go on when choosing a provider to send a bunch of sensitive data to, it'd be nice if the folks who actually know whether a provider is trustworthy would share that info.
But, since the email they sent identified the provider and the model (MiniMax, in my case) and the dates of use, it seems like it won't be secret for long what the other problem providers are. I'll assume the email to users of other providers will also include those details, so people can post about it even if OpenRouter is cagey about it in their blog post.
With this change, everything is good in the world and we can just take everyoneâs word that theyâre actually ZDR without any audits or technical enforcement again. Right?
Of course, the entire point of running an inference provider is stealing sensitive proprietary data. What are they gonna do? sue for copyright infringement?
This revealed something as delightful as a purported TEE provider having their "software TEE" root keys hacked, through "a legacy internet-facing service with management interfaces that relied on loopback access for trust and did not independently authenticate callers" (https://phala.com/posts/phala-november-preliminary-incident-...)
Did they? It says the attacker was able to create unauthorized CVMs (makes sense), able to obtain application credentials (significance unclear to me), and that they rotated their KMS root key in response (which is confusing to me because if they use hardware backed keys then I don't understand how the root key could ever be compromised).
So does this imply they aren't actually providing a proper hardware backed TEE? But that's also confusing because at least if you provision a VM you can have the hardware attest to its state.
The TEEs are real. The problem was caused by software bugs. Running a TEE app (CVM) is more than dropping a binary into TEE. Two problems remain: (1) migration, you donât want to completely lose the data if the hardware dies, and (2) upgradability, you want to patch the app when thereâs a bug without building the state from scratch. So we want to decouple the states from hardware.
Using the hardware key is painful. The solution is to introduce a separate key manager (KMS). It releases keys to the app iff. the app follows a predefined âpolicyâ (code, hardware attestation, etc). We can decouple the states from hardware based on a remote KMS.
This is rather a common design. You can find it in our system, as well as other mature frameworks like Confidential Containers.
As the system grows, a bug was introduced to the policy of how we replicate a KMSâs key material (for redundancy). The attacker exploited it with multiple sophisticated attacks, eventually being able to cheat a running KMS to hand over the keys to a fake one, like itâs a legit KMS replica. As a result, the keys protected by the KMS became at risk.
To mitigate the problem, the legacy KMS was dropped and a new KMS was deployed. We have then rotated all the keys associated with the legacy KMS. A series of protocol and code audit is ongoing to prevent future failure.
Yeah, for what I can tell (I only wasted a little time in looking into it), they say TEE TEE TEE but then oops but TEE could be compromised, it locks your data, it's inconvenient, no let's add our better software layer (where those "KMS" root keys are stored).
Ah I see, a classic case of a secure system only being as robust as the software you intentionally run on it. New crypto fails in exciting ways after all. At least it's FOSS. https://github.com/Dstack-TEE/dstack
Actually it looks like a pretty decent solution once the inevitable bugs and footguns are ironed out.
Thanks for the kind words. Any single point of failure must be handled with great care. In this case, the blockchain-based authentication is proven to be more robust than the central one. And there are a lot more to harden. Itâs a lesson we learned in a hard way.
Though frankly I would think anyone even slightly bothered about ZDR does not start their train of thought with "OK so we're going to put a middle man in"
If you're not an "enterprise", your only way to get even purported ZDR with anything close to frontier models is to use a router. The labs will not deal with individuals, and I suspect in practice it's also too hard for small businesses.
Wow I missed this story entirely. I'd seriously considered moving to OpenRouter at my next external AI renewal due to the breadth of options and pricing but this has absolutely ended that consideration.
It's clear from the breadth and variety of exposures that there's no due diligence or guardrails inside OpenRouter on who they market and how they classify their providers. Their response to this indicates a few first steps but none seem to have the force of law and i'd have expected more from a Stripe aquisition.
Very glad that my daily driver is a local model and I only use these external services sparingly.
Nice to see this move by OR.
That being said, the best ZDR agreement is the one that your company has directly with the inference provider(s). Unfortunately:
- Not everyone is large enough (or has enough $$) for that. - Ultimately youâre in the same position of having to âtrustâ that theyâre doing ZDR; even after auditing their security docs.
I'd say that's the second-best one. The most robust ZDR policy is running your own AI locally and enforcing the retention policy yourself.
Yes that's why Revolut (Europe's largest digital bank) has just started self hosting open weight models using 56 Nvidia B200 GPUs. So their 2000 engineers will use that for most things.
Expect many more companies to follow.
Thatâs not true. They switched for cost and productivity reasons.
They explicitly specified USA, EU, UK, and Israel-hosted services as acceptable as long as itâs ZDR. They said that inference hosted in China was something they didnât want.
The talk is here: https://www.youtube.com/live/akEqiXFo3GM
This right here.
ZDR is a marketing ploy. Itâs not real. Only naive people believe itâs real.
I generally trust (for example) bedrock to honor the ZDR agreement as written and to have secure infrastructure. However I don't fully trust the routing service I use to access bedrock to have rigorous security practices.
> Ultimately youâre in the same position of having to âtrustâ that theyâre doing ZDR
Could it be a moat to offer a zero-trust (based on technical impossibility) ZDR solution for corporate customers, or is everyone whoâs paranoid enough about that just going to host local models on-prem eventually?
Not an expert, I consult with medium size firms that so far have seen nice things from the AI. However, the lightly-informed C-Suite hears about the hacks and the data threats on CNBC/Bloomberg which has made them cautious. Even with their cursory understanding they are saying "can we look at these open models". N=2
Isn't that the problem themselves, in that those issues only exist if you go for open models, but outsource inference to third parties? It seems to me that there are only two safe options: either go full local and do inference on-prem/in your own cloud, or just deal with the big labs directly. Everything between these two extremes involves third parties with sketchy reputations.
I think there's definitely a market, on prem feels expensive and intimidating for many no matter what it actually costs or takes
"On prem" here just means renting a server with a GPU. (Possibly "in the cloud".)
So even more expensive.
Multiple companies are already advertising TEE-based private inference.
Yeah, like https://news.ycombinator.com/item?id=50596179
> Nice to see this move by OR.
Unfortunately, I highly disagree. I'm deeply disappointed that they don't even have the gall to call the violators out so we can be informed. Dug around for so long and I can't still find all the rule-breaking providers. It's a bunch of "trust us bro" nonsense. Name and shame should be the minimum here.
If they are going to name and shame they have to put in a lot more work first to be sure their evidence is tight enough to dissuade or dismiss defamation lawsuits, and there would still be risk and lots of drama for people too busy for that shit. It just isn't realistic to expect that in my opinion.
Its possible to be very certain of the evidence you have but not be able to prove its not been tampered with or fabricated, for example.
what move? OpenRouter is garbage, how could Stripe have thought it was a good idea to buy them? $7b for revenues that will evaporate, inevitably, all at once, in weeks!
I was affected with just one prompt by this incident, which is non issue to me because I was trying different models at that time so there was nothing serious. The troubling problem I see in this incident is the âprovidersâ in OR just casually forwarding prompts to upstream (MiniMax) instead of providing inference service them selves.
Coincidentally later on I set my OR settings to only allow some providers that I can rely on, not blindly let OR route to wherever they like.
For those who don't know this is the first thing you have to do when using openrouter, change guardrails settings from blacklist to whitelist, pick 1-2 providers you trust (do your homework, or let $agent do it for you), and if you want to go the extra mile pick the fastest/cheapest provider for you model (again do the research). The defaults can't be trusted if you care about privacy.
The first thing I did was turn on ZDR, and then looked for region hosted models. Unfortunately this is gated behind a business plan.
You can use something like Requesty (requesty.ai) where whitelisting is the default and in-region inference is available in every plan.
I got the email. The tl;dr is MiniMax (the Chinese AI lab) is actively reading your prompts, and a security researcher discovered this by putting decoy URLs into their requests which were accessed by MiniMax.
More importantly, many providers (AtlasCloud, Novita and StreamLake were named but there seems to be more implied in the blog post) have been rerouting requests to MiniMax servers on the backend and not processing their own prompts.
The obvious question is whether OpenRouter has done any auditing on their inference providers. Based on the researcher's blog post, I guess not.
https://www.dragos.com/blog/ai-supply-chain-security-canary-...
Novitaâs blog post is unreal:
https://novita.ai/blog/minimax-requests-privacy-notice/
> A routing configuration issue caused an API forwarding path intended for specific customers to apply more broadly.
> If you used MiniMax models through Novita between 2026-05-28 and 2026-11-11, your request contents may have been accessed or retained by a third party. This includes requests made directly through Novita and those routed to Novita through OpenRouter.
ârouting configuration issueâ?!? Imagine if AWS got caught storing your S3 on some other providerâs disks and they said it was a ârouting configuration errorâ! They shouldnât have software that can do this at all! The API is supposed to be for their own GPUs!
Togetherâs is weird:
> As part of routine software testing, we collected prompt samples authorized for that use and stored them in a private Hugging Face repository. The samples came from requests to the MiniMax M3, GLM-5.2, and Kimi K2.6 models, primarily on June 24, July 4, July 8, and July 14, 2026.
Did Dragos really âauthorizeâ Together to store their canary âprompt samplesâ? Somebody screwed up here.
Together.ai has made a public announcement that they are fucking clowns who cannot be trusted with anyone's data.
> During development of our public GitHub repository, xorl, an AI coding agent committed a credential that provided access to the private Hugging Face repository. Our automated scanning detected the exposed credential, and we reverted the commit.
They didn't rotate the credential they knew was exposed!?! I guess "an AI coding agent" didn't think of doing that. Also FYI kids, reverting a commit does not delete the data from the repository.
In before they claim an AI agent wrote this disclosure that included certain "errors".
https://www.together.ai/blog/security-event-prompt-leakage-f...
This is real classy coming from a neocloud with a $8.3B valuation. Shouldnât a neocloud with a bunch of Nvidia GPUs and a bunch of investment (including from Nvidia) be using those Nvidia GPUs to run inference in house? Isnât that the entire point?
Amusingly, the first time I typed âneocloudâ it got autocorrected to âbro cloudâ. I almost decided to leave it alone.
It could easily be the case that due to various reasons they could not run MiniMax at a cost lower than a negotiated wholesale API rate.
This and the way Together seems to be framing it in their disclosure post ( https://www.together.ai/blog/security-event-prompt-leakage-f... ) raises a lot of serious questions about the integrity of Together in more ways than one. Is the Together post really a fair disclosure that accounts for what Dragos identified? Their explanation is far from clear. It makes me wonder what the legal agreement was, and if the work mentioned in Together's "Polymorphic Inference" blog post, mentioning speculative decoding (using cheap models to draft) is a related practice they have routinely. Pretty worrying. I'm sure some will make this into simply blaming Chinese companies for this (not you), which wouldn't fairly account for the risks of the industries optimisation incentives. (Together's HQ is in SF AFAIK).
> Somebody screwed up here.
Yet, the basic idea of these myriads of canaries is good. I see several use cases, including in research.
You left out the best part from Togetherâs âDuring development of our public GitHub repository, xorl, an AI coding agent committed a credential that provided access to the private Hugging Face repositoryâ
So surprising that these services are vibecoded
âIâm shocked! Shocked that gambling is happening here!â
The only thing you can trust is possibly(and not always) your own hardware, your own staff and companies you can directly inspect.
Is Minimax basically an untrusted scammer at this point?
This explanation doesn't account for TogetherAI's sharing canary tokens!
As an aside, I always wondered about decision models. They are so cheap. I canât help but think that providers must be using them as honeypots for data acquisition:
You do dont find real world company documents on the web that you could train on. And companies certainly wonât upload them to you so you can improve your models, not even with incentives. But provide a model with dirt cheap input pricing and free outputs - and you gut yourself a firehose.
To think that the âlabsâ wouldnât train on this is like thinking theyâd never have trained on shadow library books because, âwell, itâs illegalâ. They did it back then and it gave them an advantage over the competition. So why wouldnât they now?
> As an aside, I always wondered about decision models. They are so cheap. I canât help but think that providers must be using them as honeypots for data acquisition:
This one is actually the simplest for me, you can run decision models on your own laptop, and you can see how cheap and fast they are, it's actually amazing how "good" (jev etc is still better) they are.
So instead of thinking of them as a "LLM" for hosting pricing, it's easier to think of them as a API as a service, it's practically "how much can we charge someone for processing in bulk for an action and to get them sticky".
You setup whatever workflow you want for the decision models and then you practically never change it, because why would you, it's dirt cheap (for you and they provider), they (the provider) is making money on your outsourcing for every API action and it's going to be consistent because it's going to be for a long time.
But nobody knows how jev is actually built , what open source decision models? All the ones available kind of jerryrig it on top of a LLM last I looked into it.
IIRC JEV is known to be built on Qwen, you simply ask it questions to that effect. No idea where I saw the thread, but someone did this exhaustively using many variations and the answer was conclusive
Decision models can be cheap because itâs a fundamentally cheaper technology than LLMs. So that part isnât suspicious at all to me.
That said, if you donât believe OpenAI when they say they donât train on your decision API inputs you shouldnât trust them regarding LLM inputs either.
You should assume that no inference provider is honoring any public commitment they have made about data retention. You should assume that 100% of what you share with an LLM over the wire is being recorded, saved, and used as training data. Any other stance is indefensible naĂŻvetĂŠ for anyone who has been around the tech industry for more than, letâs say, 5 years.
There are real exceptions to this rule, like Tinfoil.sh who cryptographically ensure they cannot retain the data. (As long as one trusts the chipmaker, I suppose.)
Of course, even with a privacy-respecting inference provider, if a modern, relentlessly proactive LLM runs in a harness with internet access, it can easily accidentally exfiltrate your data anyway.
> As long as one trusts the chipmaker
Rather you have to trust the equipment operator (which in this era of clouds isn't necessarily the vendor) not to intentionally compromise their own hardware (see https://tee.fail/). But other than that quibble yeah, tinfoil looks like an excellent option that's much less likely to leak than current alternatives.
It seems great
read anthropics tos. it says very specifically that they do not train on "inputs" and "outputs" of customer data. everything in-between is openly declared to be their legal right to train on. in terms of their ability to extract intelligence from inference there is zero difference between this and training on "inputs" and "outputs."
By the definition of how these companies see data, a simple transform to restructure the input and output would be sufficiently different to pass this bar.
exactly. the tokenization is not the input.
Thatâs why I never use a real email or real name.
The payment is where you have to provide real infos.
I just hope the payment infos is stored at a different place than prompt/conversations storage so mapping of user id to payment id is not trivial.
The payment is where you have to provide real infos.
What? I've never provided real world info when paying.
I always use Monero when paying for AI tokens. Crazy to do anything else.
More insane to use cloud provider AI. Do you care about security or privacy at all? I run my own models. Crazy to do anything else.
How do you prevent the model from exfiltrating data? Are you running it in a harness with no internet access or doing something more sophisticated?
Don't give it internet access.
Which inference providers lets you pay with Monero?
Mistral are bound by EU law
All citizens are bound by law but this doesn't mean you will have 100% confidence while walking in a dark alley at 3AM in a city you don't know.
North Koreans that could hack Mistral and sell data for crypto are not.
I fully expect them to be storing it, just the period of time they are storing it for is up for debate. Even everything older than six months ago, at this stage for me would be pure junk. If you want to train on that when I was still using soul files and barely any harness, heh, well that's your problem now!
I think tinfoil provides that(if CPU manufacturer can be trusted)
This reminds again that all "promises" of companies are a lie and you better not trust their "privacy policies".
They are not all lies, but these are definitely circumstances in which case lies should be expected because it is possible to profit significantly by making them. Lots of new companies springing up over night making promises they won't steal the valuable data passing through their hands.
The number of providers they have added is absurd, there have to be some that have almost no compute and just upstream everything. So stealing is their only possible business model.
openrouter.ai hasn't really promised that they do any vetting or due diligence, as far as I know. This whole space has had the scammy crypto vibe from the beginning.
ZDR through AWS Bedrock is still safe right? AFAIK Bedrock self hosts all models they provide. Even proprietary weight models like Claude are run within AWS network.
Assuming you trust AWS is not, say tapped by NSA. Of course, if your data is on AWS already, this makes no difference.
Yes, however some models donât allow ZDR IIRC. I think Fable requires enabling data retention.
Bedrock provides an incredibly limited and old set of models. Anthropic and OpenAI models are not hosted by AWS they are passed through.
Also saw Prompt leakage for requests incident from Together AI
https://www.together.ai/blog/security-event-prompt-leakage-f...
though this is not similar as ZDR, it's crazy to see so many incidents in last few months.
Is there any information on the three companies that violated ZDR? I could only find mention of one of those affected providers in the original article (that OR links in their post): Lark. Trying to figure out if I ended up routing to one of the affected "ZDR" endpoints.
I don't understand why these names are being coyly tiptoed around as if they're trying to protect the privacy of these companies blatantly lying about ZDR. Why isn't OR calling out the actual providers they off-boarded by name? Why is OR basically protecting them under a cover of anonymity?
Lost a lot of trust in OR that's for sure, in-spite of some other positive reception I see here for this move.
Edit: Wording update.
They state they notified customers who used any of the problem models via email (I got an email about it), which answers your question of whether you ended up routing to one of those endpoints. They apparently know who was routed to all those providers, so you don't have to guess.
That still leaves the desire of people to make informed decisions about what providers to use unresolved. If reputation is the only thing we have to go on when choosing a provider to send a bunch of sensitive data to, it'd be nice if the folks who actually know whether a provider is trustworthy would share that info.
But, since the email they sent identified the provider and the model (MiniMax, in my case) and the dates of use, it seems like it won't be secret for long what the other problem providers are. I'll assume the email to users of other providers will also include those details, so people can post about it even if OpenRouter is cagey about it in their blog post.
> OpenRouter is cagey about it in their blog post.
That's exactly what's confusing me though? Why is OR basically covering for these providers by not naming them?
So they don't get sued, you bet this was reviewed by lawyers.
Also, naming names makes it real and bigger than this reads now with its vagaries.
Here:
https://x.com/nateberkopec/status/2123992018850595164?a=UMOK...
> Just got the email from Openrouter and now the story is confirmed
> Atlascloud, Novita (who I know was flagged US with ZDR on openrouter) and Streamlake were⌠oops! Forwarding your data BACK TO MINIMAX IN CHINA
They just don't care. That is the reason.
Given how trust based the LLM world has to be right now, there should be zero tolerance for anyone doing stupid shit like this
Is there any information on the three companies that violated ZDR?
I have my harness set to show me "diffs" of the catalog when I start a new session.
There was a mass-downgrade to non-ZDR a few days ago. At the time I wondered why.
It was all older (not latest release) Chinese models being run by second-party inference providers.
GMI Cloud was removed as a model provider about the time the news went out.
...draw your own conclusions....
Model routing was always insane. Until there is attestation from the providers and encryption of the session the temptation to peek will be too great.
I hope this time this stays on the frontpage, every LLM user should be aware of it.
There's also a good article at https://arstechnica.com/security/2026/11/a-leaky-ai-supply-c... , and a few earlier comments at https://news.ycombinator.com/item?id=50582794 , https://news.ycombinator.com/item?id=50585962 and https://news.ycombinator.com/item?id=50562582 .
As usual, if you really give a shit about your data, donât give it to someone else.
With this change, everything is good in the world and we can just take everyoneâs word that theyâre actually ZDR without any audits or technical enforcement again. Right?
Zero data retention, give or take 125 endpoints.
Of course, the entire point of running an inference provider is stealing sensitive proprietary data. What are they gonna do? sue for copyright infringement?
This revealed something as delightful as a purported TEE provider having their "software TEE" root keys hacked, through "a legacy internet-facing service with management interfaces that relied on loopback access for trust and did not independently authenticate callers" (https://phala.com/posts/phala-november-preliminary-incident-...)
> having their "software TEE" root keys hacked
Did they? It says the attacker was able to create unauthorized CVMs (makes sense), able to obtain application credentials (significance unclear to me), and that they rotated their KMS root key in response (which is confusing to me because if they use hardware backed keys then I don't understand how the root key could ever be compromised).
So does this imply they aren't actually providing a proper hardware backed TEE? But that's also confusing because at least if you provision a VM you can have the hardware attest to its state.
Cofounder of Phala here. Happy to explain it.
The TEEs are real. The problem was caused by software bugs. Running a TEE app (CVM) is more than dropping a binary into TEE. Two problems remain: (1) migration, you donât want to completely lose the data if the hardware dies, and (2) upgradability, you want to patch the app when thereâs a bug without building the state from scratch. So we want to decouple the states from hardware.
Using the hardware key is painful. The solution is to introduce a separate key manager (KMS). It releases keys to the app iff. the app follows a predefined âpolicyâ (code, hardware attestation, etc). We can decouple the states from hardware based on a remote KMS.
This is rather a common design. You can find it in our system, as well as other mature frameworks like Confidential Containers.
As the system grows, a bug was introduced to the policy of how we replicate a KMSâs key material (for redundancy). The attacker exploited it with multiple sophisticated attacks, eventually being able to cheat a running KMS to hand over the keys to a fake one, like itâs a legit KMS replica. As a result, the keys protected by the KMS became at risk.
To mitigate the problem, the legacy KMS was dropped and a new KMS was deployed. We have then rotated all the keys associated with the legacy KMS. A series of protocol and code audit is ongoing to prevent future failure.
Yeah, for what I can tell (I only wasted a little time in looking into it), they say TEE TEE TEE but then oops but TEE could be compromised, it locks your data, it's inconvenient, no let's add our better software layer (where those "KMS" root keys are stored).
(https://docs.phala.com/dstack/design-documents/decentralized... , https://docs.phala.com/dstack/design-documents/key-managemen...)
Ah I see, a classic case of a secure system only being as robust as the software you intentionally run on it. New crypto fails in exciting ways after all. At least it's FOSS. https://github.com/Dstack-TEE/dstack
Actually it looks like a pretty decent solution once the inevitable bugs and footguns are ironed out.
Thanks for the kind words. Any single point of failure must be handled with great care. In this case, the blockchain-based authentication is proven to be more robust than the central one. And there are a lot more to harden. Itâs a lesson we learned in a hard way.
Solid move.
Though frankly I would think anyone even slightly bothered about ZDR does not start their train of thought with "OK so we're going to put a middle man in"
If you're not an "enterprise", your only way to get even purported ZDR with anything close to frontier models is to use a router. The labs will not deal with individuals, and I suspect in practice it's also too hard for small businesses.