+− THE DAILY DIFFdev & AI news
REVERT

OpenRouter removed 125 endpoints from zero data retention: what researchers, Novita, Phala and Together say

Researchers at the security company Dragos report that they sent about two million prompts between July and November 2026, each carrying a unique fake credential and secret link, and watched where the secrets surfaced.

Want to see our stories first on Google? Add us to your preferred sources — it helps us grow.

Researchers at the security company Dragos report that they sent about two million prompts between July and November 2026, each carrying a unique fake credential and secret link, and watched where the secrets surfaced. They report evidence that data from at least three providers was exposed, that some requests marked zero data retention were affected, and that some providers did not honor privacy settings. On November 22, OpenRouter said the findings showed no breach of OpenRouter, that it off-boarded three providers, removed 125 endpoints from zero data retention eligibility (including some only as a precaution) and notified potentially affected customers, and that it is adapting the researchers' tests to run on an ongoing basis and requiring assurances before restoring eligibility. Its post does not name the three removed providers; it links notices from Novita, Phala and Together. Novita says some MiniMax requests between 2026-05-28 and 2026-11-11 were forwarded to MiniMax's own API because of a routing configuration issue, and suspended the affected models. Phala's preliminary report says an attacker took control of an inference machine and that requests to one DeepSeek model routed there from October 21 to November 8 should be treated as exposed, while the researcher's canaries were misused only from October 21 and 22. Together says an AI coding agent committed a credential for a private repository holding prompt samples it had kept for testing; an unauthorized party downloaded them and tested credentials found inside, five of them the researchers' canaries. The episode also corrects its own Saturday status on the Riemann Hypothesis Lean proof: the Lean and Mathlib communities certified the Lean statement and proof at 13:55 UTC on November 21, about ninety minutes before the earlier episode went public, while Andrew Sutherland's MIT page, as read at 21:58 GMT+7 on November 22, still listed the full comparator run and Lean's own kernel replay as running. The verdict is REVERT: revert the trust, keep the router.

Read the written edition (English) ↗

What this video covers

  • Dragos reports that it sent about two million prompts with unique fake secrets from July to November 2026 and that data from at least three providers was exposed; some stolen requests were marked zero data retention (Dragos, November 19).
  • OpenRouter says it off-boarded three providers, removed 125 endpoints from zero data retention eligibility and notified potentially affected customers; its post does not name the three removed providers (November 22).
  • OpenRouter's documentation defines zero data retention as no routine storage of prompts or responses after processing, with an exception for abuse screening; the label comes from OpenRouter working out each host's policy with the host.
  • Novita says some MiniMax requests, including requests routed through OpenRouter, were forwarded to MiniMax's own API between 2026-05-28 and 2026-11-11 because of a routing configuration issue.
  • Phala says requests to one DeepSeek model that OpenRouter routed to it from October 21 to November 8 should be treated as exposed; the researcher's canaries were misused only from October 21 and 22.
  • Together says prompt samples it kept for testing in a private repository were downloaded after an AI coding agent committed a credential; this is a different failure from a broken zero retention flag.
  • The researchers say a privacy flag only binds the first hop; they suggest picking upstream providers yourself and planting your own fake credential as a canary.
  • Riemann update: the Lean and Mathlib communities certified the Lean statement and proof on November 21; mathematicians say digesting the argument will take time, and the MIT page still listed the full comparator run as running as of November 22, 21:58 GMT+7.
  • Visual credits: portrait of HD Moore (2018), own work, CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), cropped with the background removed and shared under the same licence; it identifies the author of the Dragos report only, with 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. Meme templates via imgflip.

Questions and answers

What does zero data retention promise on a router like OpenRouter?

OpenRouter's documentation says zero data retention means no routine storage of prompts or responses after processing, apart from an exception for screening child sexual abuse material. When enforced, in account settings, a guardrail or per request, requests go only to endpoints whose host says it keeps nothing. The documentation says OpenRouter works with providers to understand each of their data policies. It also says enforcement covers provider routing, not plugins and tools such as web search, which have their own policies.

Watch this part 0:31 · OpenRouter documentation

How did researchers test whether providers honor zero data retention?

Dragos reports that it sent about two million prompts between July and November 2026, each carrying a unique fake credential and secret web link that exist nowhere else. Any later use of one shows which prompt leaked and which route and privacy setting carried it. The team also built a traceroute for AI inference to find hidden hops. Dragos says some stolen requests were marked zero data retention and that a privacy flag only binds the first hop.

Watch this part 0:46 · Dragos (HD Moore)

What did Novita, Phala and Together each say happened?

Novita says some MiniMax requests between May 28 and November 11 were forwarded to MiniMax's own API because of a routing configuration issue, and it suspended the affected models. Phala's preliminary report says an attacker took control of an inference machine, and that requests to one DeepSeek model routed there from October 21 to November 8 should be treated as exposed. Together says prompt samples it kept for testing were downloaded after an AI coding agent committed a credential.

Watch this part 1:18 · Novita · Phala · Together AI

What did OpenRouter do after the Dragos report, and what can a builder check?

OpenRouter says it off-boarded three providers, removed 125 endpoints from zero data retention eligibility, some only as a precaution, and notified potentially affected customers. Its post does not name the three providers. It is adapting the researchers' tests to run on an ongoing basis. Builders can review their OpenRouter privacy settings and guardrails, consult the public endpoint list, pick upstream providers themselves as the researchers suggest, and rotate secrets that were in affected requests, as Novita and Phala advise.

Watch this part 2:00 · OpenRouter · Novita · OpenRouter documentation

What was The Daily Diff's verdict on the zero data retention story?

The Daily Diff's verdict was REVERT: revert the trust, keep the router. A zero data retention flag is a promise until someone tests it where prompts actually go, and until OpenRouter's continuous tests exist, the builder is the tester. The verdict applies to relying on the label without checking, not to OpenRouter's response, which the episode describes as the right set of steps announced but not yet shipped.

Watch this part 4:47 · OpenRouter · Dragos (HD Moore)

Transcript

What did researchers hide in the prompts?

0:00 You switched on zero data retention, so you assume nobody keeps your prompt. Researchers hid fake secrets in nearly two million prompts, and strangers later tried secrets from prompts marked zero retention. In this video, how does a zero retention promise fail? What can you check from your side? And what should you change this week? One notice says an AI coding agent committed the credential that started it. I'll tell you which notice, and what happened next,

0:25 before the verdict. It's Sunday, November twenty-second, and this is The Daily Diff. OpenRouter is a router, one API in front of many models and the companies that

What does zero data retention promise?

0:34 host them. Its zero data retention setting sends your request only to hosts that say they keep nothing after processing. That label comes from OpenRouter working out each host's policy with the host. Security company Dragos tested the promise.

How did researchers catch it?

0:47 From July to November its team sent nearly two million prompts, each with its own fake credential and secret web link. If anyone ever uses one, you know exactly which prompt leaked, and which route and privacy setting carried it. The strangest result is a chat message. Twenty-four days after one test prompt went out with the zero retention flag, the researchers say its secret link was pasted into a company chat in China and opened within two minutes from six devices.

1:14 They conclude someone was reading a stored copy of the prompt.

What did the providers admit?

1:18 Novita says some MiniMax requests were forwarded to MiniMax's own API instead of running on Novita's servers, from May twenty-eighth to November eleventh, including requests routed through OpenRouter. It calls that a routing configuration issue, and it suspended the affected models. Phala's failure is different. An attacker chained a legacy internet-facing service with management interfaces that trusted local callers, then took control of an inference machine on a new

1:43 GPU server serving production traffic without full hardening. Phala says requests to one DeepSeek model that OpenRouter routed there from October twenty-first to November eighth should be treated as exposed, though only the researcher's canaries from the first two days were later misused. OpenRouter says none of this was a breach of OpenRouter.

What did OpenRouter do?

2:02 It off-boarded three providers, removed one hundred twenty-five endpoints from zero retention eligibility, some only as a precaution, and notified customers it thinks were affected. It also says it will adapt the researchers' tests to run on an ongoing basis, and demand assurances before restoring zero retention status to any removed endpoint. Three providers left the list, and the post does not say which. The post links three notices, says not every affected provider has published

2:28 one, and Hacker News commenters want names. The researchers say a privacy flag only binds the first hop.

What does the flag actually cover?

2:35 Their traceroute for AI inference found extra layers nobody advertised, and a handful of traces sent their test image to a company that was not on the advertised path. What can you check yourself?

What should you check this week?

2:45 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, and planting your own free fake credential in what your agents send, treating any use of it as an incident. Novita and Phala both say rotate any secrets in affected requests. Now the notice I promised.

Which notice names a coding agent?

3:07 Together says it kept prompt samples for testing in a private Hugging Face repository. While building a public repository, an AI coding agent committed a credential for the private one. The commit was reverted, but in the window an unauthorized party downloaded the samples and tested credentials found inside. Five were the researchers' fake ones. The samples were kept on purpose, so this is a retention failure of a different kind. Together now makes zero retention the default for new accounts.

Did the Riemann check finish?

3:33 One more thing, because yesterday's story moved. In yesterday's video I told you the comparator run on the Riemann proof was still going on Saturday afternoon. The Lean and Mathlib communities had already posted a joint statement about ninety minutes before it went public, saying all four of Kevin Buzzard's criteria are satisfied. So that status was out of date when you saw it. The statement says the theorem in the proof is exactly Mathlib's own statement,

3:56 confirmed by two independent comparison tools. Several checkers then accepted the exported proof. Two of them come with a formal proof, relative to set theory, that they cannot accept a false proof, and the checks ran on an inert copy that malicious code could not touch. On Saturday Scott Aaronson added an update to his post saying the proof is correct. Terence Tao wrote on Friday that he imagines the digestion process will take some time.

4:20 Ravi Vakil, who posted the rival human draft, says he makes no claim as to the correctness of its proof strategy. And when I read Andrew Sutherland's page just before ten tonight, Bangkok time, it said it was last updated at nine ten tonight. It still listed the full comparator run and Lean's own kernel replay as running, with its statement check already passed. If you'd rather read this than hear me say it, the diff lands in your inbox every morning, free at the daily diff dot dev, link below.

4:46 So today's verdict on the zero retention story is REVERT.

What is today's verdict?

4:50 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. Subscribe, hit the bell, and tell me in the comments if you'd have stamped it differently. And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

Sources

  1. Data Processing Incidents on Inference ProvidersOpenRouter
  2. Three breaches, endless leaks: Your AI supply chain needs canariesDragos (HD Moore)
  3. Privacy Notice Regarding MiniMax Requests on NovitaNovita
  4. Phala November Preliminary Incident ReportPhala
  5. Security event: Prompt leakage for requests in June and July of 2026Together AI
  6. Zero Data RetentionOpenRouter documentation
  7. Data Processing Incidents on Inference Providers (discussion)Hacker News
  8. Lean verification of the Riemann HypothesisLean and Mathlib communities
  9. Verification files for the Lean proof of the Riemann HypothesisAndrew Sutherland, MIT
  10. The stochastic parrots have apparently proved the Riemann HypothesisScott Aaronson, Shtetl-Optimized
  11. Unforced Euler singularity, unforced Navier-Stokes singularity, and the Generalized Riemann HypothesisTerence Tao, What's new
  12. Hurried Announcement: Alessandro Masullo GRHRavi Vakil, Math 216

Niko from Axrisi

Nikoloz Turazashvili · Vexrail · Axrisi

LinkedIn · dev.to · GitHub · X · Substack

Related videos