OpenRouter says it "monitors every endpoint, from every provider" to know its data retention policy.
This month its users still learned about 2 leaks from outside researchers:
Nov 10: Phala, a host OpenRouter routes to, said an attacker had control of the server answering OpenRouter's DeepSeek V4 Flash requests from Oct 21 to Nov 8.
Today: OpenRouter emailed users that prompts sent to MiniMax models "may have been read." A researcher planted unique URLs in them, and someone opened those URLs later. Some of those prompts went to endpoints labeled Zero Data Retention.
OpenRouter says the issue sits with MiniMax and the providers that forwarded requests to it. The ZDR label came off those endpoints after the researcher's test. So what was the label checking?
We've been saying ZDR is a lie for a year. A ZDR label is a provider's terms of service that was passed along to you. You can't check a ToS from the outside.
And this isn't a Chinese-model problem. Send a prompt to any closed model and you have no idea what code ran, what weights ran, or who kept a copy.
TrustedRouter's confidential routes run on hardware that proves which code and which weights handled your prompt. If a provider can't prove it, the request gets refused.
If you got OpenRouter's email today, moving to our confidential routes is one line of code:
https://t.co/uSTC0fLjbm
got the openrouter email?
1/ rotate any keys that were in those prompts
2/ check which provider actually ran them
3/ for anything sensitive, use a route that proves what's running before it takes your prompt
4/ or switch to https://t.co/xv5duzG5bJ's confidential routes. that's what we built it for
OpenRouter: prompts sent to MiniMax models "may have been read."
A researcher planted unique URLs in them, and someone opened those URLs later. Some of those prompts went to endpoints labeled Zero Data Retention.
The need for trust via encyption is clear. CHeckout @TrustedRouter
6/ every one of these had βa policyβ but none had proof. on a confidential route the chip proves the exact code and weights before your prompt goes in. if it can't prove it the request fails
@jsrailton and the fix already exists. the web went through this with https. inference can run inside sealed hardware that proves what code touches your prompt. it just isn't the default yet
got the openrouter email?
1/ rotate any keys that were in those prompts
2/ check which provider actually ran them
3/ for anything sensitive, use a route that proves what's running before it takes your prompt
4/ or switch to https://t.co/xv5duzG5bJ's confidential routes. that's what we built it for
step by step on what to check and rotate after the openrouter email: https://t.co/pxREmG1FU3
OpenRouter says it "monitors every endpoint, from every provider" to know its data retention policy.
This month its users still learned about 2 leaks from outside researchers:
Nov 10: Phala, a host OpenRouter routes to, said an attacker had control of the server answering OpenRouter's DeepSeek V4 Flash requests from Oct 21 to Nov 8.
Today: OpenRouter emailed users that prompts sent to MiniMax models "may have been read." A researcher planted unique URLs in them, and someone opened those URLs later. Some of those prompts went to endpoints labeled Zero Data Retention.
OpenRouter says the issue sits with MiniMax and the providers that forwarded requests to it. The ZDR label came off those endpoints after the researcher's test. So what was the label checking?
We've been saying ZDR is a lie for a year. A ZDR label is a provider's terms of service that was passed along to you. You can't check a ToS from the outside.
And this isn't a Chinese-model problem. Send a prompt to any closed model and you have no idea what code ran, what weights ran, or who kept a copy.
TrustedRouter's confidential routes run on hardware that proves which code and which weights handled your prompt. If a provider can't prove it, the request gets refused.
If you got OpenRouter's email today, moving to our confidential routes is one line of code:
https://t.co/uSTC0fLjbm
Providersβ data-handling claims have mostly relied on self-attestation. This isn't going to be sufficient.
Just as continuous benchmarking and monitoring transformed the quality and reliability of inference both on and out of OpenRouter, we expect technical validation, auditing, and continuous testing to do the same for data policy enforcement.
Encrypted TEE (Trusted Execution Environment) inference will be a great way for providers to accelerate this.
We will keep raising the bar, and providers that want to be on OpenRouter will need to meet it.