Watch the video, transcript and sources
+ OpenRouter: 125 endpoints removed from zero data retention eligibility
+ Novita, Phala and Together each published a notice
- Dragos: strangers later tried secrets from prompts marked zero retentionYou switched on zero data retention, so you assume nobody keeps your prompt. Researchers at the security company Dragos hid fake secrets in nearly two million prompts between July and November, and say strangers later tried secrets from prompts marked zero retention. This is what each party says, what the flag covers, and what you can check this week. Nothing below comes from my own testing.
What the flag promises. OpenRouter is a router: one API in front of many models and the companies that host them. Its zero data retention setting sends your request only to endpoints whose host says it keeps nothing after processing. OpenRouter's docs say that label comes from working out each host's policy with the host.
How researchers caught it. Dragos' HD Moore describes canary tokens: a secret that exists in exactly one place, so any use tells you that place was exposed. Each test prompt carried its own fake credential and secret link, tied back to the route and privacy setting that carried it. One link from a test prompt was, the researchers say, pasted into a company chat in China twenty-four days later and opened within two minutes from three desktops and three phones; they conclude someone was reading a stored copy of the prompt.
What the providers say. Novita says some MiniMax requests were forwarded to MiniMax's own API between May 28 and November 11, including requests routed through OpenRouter, because of "a routing configuration issue", and that it suspended the affected models. Phala says an attacker chained a legacy internet-facing service with management interfaces that trusted local callers and took control of an inference machine on a new GPU server; requests to one DeepSeek model that OpenRouter routed there from October 21 to November 8 should be treated as exposed, although only the researcher's canaries from October 21 and 22 were later misused. Together says an AI coding agent committed a credential for a private Hugging Face repository holding prompt samples it kept for testing; an unauthorized party downloaded the samples during the exposure window and tested credentials found inside, five of them the researchers' fake ones. Those samples were kept on purpose, so that is a different failure from a broken flag.
What OpenRouter did. It says none of this was a breach of OpenRouter, that it off-boarded three providers, removed 125 endpoints from zero data retention eligibility (some only as a precaution) and notified potentially affected customers. It says it is adapting the researchers' tests to run on an ongoing basis and will require assurances before restoring eligibility. Its post does not name the three providers it removed, and says not every affected provider has published a notice. The three linked notices are not that list.
What the flag covers. The researchers say a privacy flag only binds the first hop. Their traceroute for AI inference found layers the advertised route never mentions.
What to check. OpenRouter's docs say settings and guardrails decide whether zero retention is enforced, and its list of qualifying endpoints is public. The researchers suggest picking your upstream providers yourself, planting your own free fake credential in what your agents send, and treating any use of it as an incident. Novita and Phala both say to rotate secrets that were in affected requests.
One more thing: yesterday's Riemann check. Our Saturday video said the comparator run was still going on Saturday afternoon. The Lean and Mathlib communities had posted a joint statement at 13:55 UTC, about ninety minutes before that video went public, saying all four of Kevin Buzzard's criteria are satisfied. So that status was out of date. Scott Aaronson added an update to his post saying the proof is correct; Terence Tao wrote that he imagines the digestion process will take some time; Ravi Vakil says he makes no claim as to the correctness of the rival human draft's proof strategy. As read at 21:58 GMT+7 on Sunday, Andrew Sutherland's MIT page was last updated at 21:10 GMT+7 and still listed the full comparator run and Lean's own kernel replay as running, with its statement check already passed.
Verdict: REVERT. I'd revert the trust and keep the router. A zero retention flag is a promise until someone tests it where your prompts actually go, and until OpenRouter's continuous tests exist, that someone is you.
Sources
https://openrouter.ai/blog/announcements/data-processing-incidents-on-inference-providers/
https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens
https://novita.ai/blog/minimax-requests-privacy-notice/
https://phala.com/posts/phala-november-preliminary-incident-report
https://www.together.ai/blog/security-event-prompt-leakage-for-requests-in-june-and-july-of-2026
https://openrouter.ai/docs/guides/features/zdr
https://news.ycombinator.com/item?id=50595554
https://leanprover-community.github.io/blog/posts/riemann-hypothesis/
https://math.mit.edu/~drew/grh/
https://scottaaronson.blog/?p=10344
https://terrytao.wordpress.com/2026/11/20/unforced-euler-singularity-unforced-navier-stokes-singularity-and-the-generalized-riemann-hypothesis/
https://math216.wordpress.com/2026/11/20/hurried-announcement-alessandro-masullo-grh/
And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.
YouTube · Newsletter · thedailydiff.dev · forward this to whoever trusts a flag without asking where the prompt went.
Credits: Portrait of HD Moore (2018), own work, CC BY-SA 4.0 (licence, Commons file); cropped, background removed, shared under the same licence; identifies the author of the Dragos report only, no endorsement, quote or reaction. Silent 1.4-second quotation of a Homer Simpson hedge GIF via GIPHY; the owner keeps all rights and nobody in this story reacted. "Always Has Been" and "Gru's Plan" meme templates via imgflip.

