OpenRouter zero data retention: what the Dragos canary test and three provider notices say
OpenRouter said on November 22, 2026 that it off-boarded three inference providers, removed 125 endpoints from zero data retention eligibility and notified customers it considers potentially affected. The trigger was res
OpenRouter said on November 22, 2026 that it off-boarded three inference providers, removed 125 endpoints from zero data retention eligibility and notified customers it considers potentially affected. The trigger was research from the security company Dragos, which hid unique fake secrets in about two million prompts between July and November and watched where they surfaced. This post lays out what each party actually says, with dates, and which of the checks are in your hands. Everything here is attributed: I ran no test and read no provider's logs.
TL;DR
- Dragos reports that data from at least three AI providers was exposed, that some stolen requests were marked zero data retention (ZDR), and that some providers did not honor privacy settings (Dragos, Nov 19).
- OpenRouter says the findings showed no breach of OpenRouter, that it off-boarded three providers, removed 125 endpoints from ZDR eligibility, some only as a precaution, and notified potentially affected customers (OpenRouter, Nov 22).
- OpenRouter's post does not name the three removed providers. It links notices from Novita, Phala and Together, which are not that list.
- Together's incident (prompt samples kept for testing) is a different failure from a broken ZDR flag.
- The researchers say a privacy flag only binds the first hop.
What zero data retention promises
OpenRouter is a router: one API in front of many models and the companies that host them. Its documentation says ZDR "means no routine storage of prompts or responses after processing", apart from an exception for screening child sexual abuse material (ZDR docs). You can enforce it in account privacy settings, per model group, in a guardrail, or per request with a provider.zdr field. When it is enforced, your request goes only to endpoints classified as ZDR.
The classification comes from OpenRouter working out each provider's data policy with the provider, and in some cases special agreements. When a policy cannot be established, the docs say OpenRouter assumes the endpoint both retains and trains on data. The docs also say enforcement covers provider routing only: plugins and tools such as web search may be run by third parties with their own policies.
How Dragos tested the promise
A canary token is a secret that exists in exactly one place, so any use of it tells you that place has been exposed. Dragos' team, led by Chief Research Officer HD Moore, began in July with free Thinkst Canarytokens and then built a service that generates a unique AWS-style key, hostnames and a URL for every prompt, with the key tied to an empty, fully logged account. Each hit maps back to the exact prompt, model, router and privacy setting that carried it. Over five months it submitted approximately two million prompts.
To find which hops handled a request, Dragos built what it calls a traceroute for AI inference: an image URL in the request reveals every system that fetches it, and deliberately slightly broken requests make each router wrap the error from the hop behind it, which exposes nested signatures. It reports tracing about two thousand routes, several hundred including a system it could not identify, and a handful sending the image to a company that was not on the advertised path at all.
Its three findings, in its words and numbers: attackers actively stealing data from three hosting providers, with some stolen requests in two of those cases marked ZDR; marketplace resellers who could view and retain prompts using standard logging; and providers that appeared not to honor privacy settings. In one example, the researchers say a link from a test prompt was pasted into a company chat in China twenty-four days later and opened within two minutes from three desktops and three phones; they conclude that whoever shared it was reading a stored copy of the full prompt. Dragos does not name the three compromised providers or the Chinese lab.
What the three provider notices say
Novita (November 22). Some MiniMax requests on Novita between 2026-05-28 and 2026-11-11 were forwarded to MiniMax's official API instead of being processed on Novita-hosted infrastructure, including requests routed through OpenRouter. Novita says "a routing configuration issue caused an API forwarding path intended for specific customers to apply more broadly". The researcher observed accesses to URLs and hostnames embedded in test requests; no use of the AWS test key markers was observed, and Novita says the pattern suggests automated content review may have been involved, which it is following up with MiniMax. It suspended the affected models and recommends rotating live credentials as a precaution.
Phala (preliminary report, November 10). An attacker chained a legacy internet-facing service with management interfaces that relied on loopback access for trust, and later took control of an inference virtual machine on a new GPU server that was serving production traffic without full hardening. Phala says requests to DeepSeek V4 Flash 0731 that OpenRouter routed to it between October 21, 00:00 UTC and November 8, 05:35 UTC should be treated as exposed. Only the researcher's canary credentials from requests sent on October 21 and 22 were later misused; Phala keeps the wider window because the tests sampled requests. It contained the incident on November 8 and promises a full root cause analysis.
Together AI (November 21). Together says it kept prompt samples, authorized for that use, in a private Hugging Face repository for routine testing, from requests to three open models, primarily on June 24, July 4, July 8 and July 14. An AI coding agent committed a credential for that repository to a public GitHub repository; its scanning caught it and the commit was reverted. During the exposure window an unauthorized party used it to download the samples and tested credentials found inside; five were Dragos canaries. Together engaged a forensics firm and law enforcement, says it found no further leakage, and made ZDR the default for new accounts.
What OpenRouter changed
OpenRouter lists three actions taken (three providers off-boarded, 125 endpoints removed from ZDR eligibility, customers notified) and three underway: adapting the researchers' testing approach to run on an ongoing basis, technical verification of network paths where requests may be forwarded to third parties, and requiring assurances before restoring ZDR eligibility to any removed endpoint. It says not all affected providers have posted publicly.
What you can check this week
Every item is attributed:
- Review OpenRouter privacy settings and guardrails: the docs say they decide whether ZDR is enforced.
- The list of qualifying endpoints is public at OpenRouter's documented endpoint list, so you can compare it with the endpoints you used.
- The researchers suggest picking upstream providers yourself instead of accepting the default pool, and verifying that your privacy settings change where requests go.
- Novita and Phala both say to rotate secrets that were in affected requests.
- The researchers suggest planting your own free fake credential, such as a Thinkst Canarytoken, in the context your agents already send, and treating any use of it as an incident.
Where this leaves the label
A ZDR flag is OpenRouter's classification of a host's policy. The researchers' point is that the classification binds the first hop, and that checking the actual path takes a test. OpenRouter says it is building one. I stamped the story REVERT: revert the trust, keep the router.
One more thing: yesterday's Riemann check
Our November 21 episode said the official comparator run on the Lean proof of the Riemann Hypothesis was still going on Saturday afternoon. The Lean and Mathlib communities had posted a joint statement at 13:55 UTC that day, about ninety minutes before that video went public, saying all four of Kevin Buzzard's criteria are satisfied, so our status was out of date. Scott Aaronson updated his post to say the proof is correct, Terence Tao wrote that he imagines the digestion process will take some time, and Ravi Vakil, who posted a rival human draft, says he makes no claim as to the correctness of its proof strategy. As read at 21:58 GMT+7 on November 22, 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 (MIT page). That page keeps changing.
Sources
- OpenRouter: Data Processing Incidents on Inference Providers
- Dragos: Three breaches, endless leaks
- Novita notice
- Phala preliminary incident report
- Together AI security event
- OpenRouter ZDR documentation
- Hacker News discussion
- Lean and Mathlib communities statement
- Scott Aaronson ยท Terence Tao ยท Ravi Vakil
Visual credits: portrait of HD Moore (2018), own work, CC 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.
This article expands on an episode of **The Daily Diff, a five-minute daily video on what shipped and what broke in tech.
Watch the episode ยท Subscribe on YouTube ยท the written diff lands in your inbox every morning at thedailydiff.dev.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ full credit and traffic to the original publisher.