Independent/Reader-funded/Infrastructure, not tokens
DeAINEWS

AI you control — open models, private inference, and the networks that run them.

Provider Policy & Trust

OpenRouter off-boards three providers, strips 125 ZDR endpoints

OpenRouter off-boarded three providers and stripped 125 endpoints of Zero Data Retention status after canary research traced leaked prompts to MiniMax routes.

DeAI is powered by Morpheus (mor.org). We cover competing providers on the same terms — see our methodology.

A rack server with patch cables and a glowing amber status LED in a datacenter corridor, depicting OpenRouter stripping Zero Data Retention labels from provider endpoints after canary research. Illustration: DeAI
A rack server with patch cables and a glowing amber status LED in a datacenter corridor, depicting OpenRouter stripping Zero Data Retention labels from provider endpoints after canary research. Illustration: DeAI

OpenRouter has taken its first marketplace-enforcement actions after a canary-token study caught prompt leakage in the AI supply chain: the routing platform says it has off-boarded three providers, stripped 125 endpoints of Zero Data Retention eligibility, and notified every customer whose data may have been exposed through routes tied to China-based MiniMax.

Key facts

  • OpenRouter off-boarded three providers and removed 125 endpoints from Zero Data Retention (ZDR) eligibility, per its November 22, 2026 incident post.
  • Dragos chief research officer HD Moore's team seeded roughly two million prompts with unique URLs and canary credentials over five months, July through November 2026.
  • OpenRouter's November 21 customer email said test URLs embedded in MiniMax prompts were fetched 40 to 130 hours after inference, between July and October 2026.
  • Three resellers — AtlasCloud, Novita, and StreamLake — told OpenRouter they forward MiniMax listings to MiniMax itself, per the customer email.
  • In the largest breach Dragos documented, a couple hundred baited AWS keys sent in single requests were all stolen and driven tens of thousands of API calls against the researcher's test account.
  • Three providers have posted public incident write-ups so far — Novita, Phala, and Together — with OpenRouter noting not all affected providers have done so.

What happened

The sequence started with the research. On November 19, Dragos published a study in which HD Moore's team sent nearly two million AI prompts containing unique, trackable secrets — URLs and AWS credentials no legitimate system would ever use — to widely used AI services, then watched where those secrets surfaced. Over the five-month window the team traced about two thousand routes and reported three categories of findings: evidence of compromises at three providers, marketplace resellers viewing and retaining customer prompts with standard logging tools, and providers failing to honor published privacy settings. In one case, prompts sent to US-based providers were transmitted to Chinese AI organizations, and one seeded link eventually appeared in an internal company chat in China.

The most direct finding involved outright theft. Baited AWS keys sent in single requests during late-October windows were stolen in full, including keys belonging to requests the provider rejected before a model ever ran — which, Moore wrote, indicates the attacker was copying traffic at the provider's front door. The stolen keys were tested, used to enumerate S3 buckets and search Secrets Manager and Parameter Store, and driven through tens of thousands of AWS API calls, including attempts to enable models in AWS Bedrock and add attacker-controlled IAM users. Some of the stolen requests were marked ZDR. Ars Technica's security coverage framed the study as evidence of a leaky AI supply chain.

Two days later, OpenRouter began emailing paying customers. The November 21 email, excerpts of which developers posted publicly, told each account how many requests it had sent to MiniMax models, which upstream hosts carried them, and over which date range. OpenRouter said a third party fetched test URLs embedded in those prompts 40 to 130 hours after the requests completed, a lag consistent with stored prompts being read later rather than real-time filtering. It said the pattern held whether MiniMax answered directly or one of three named resellers — AtlasCloud, Novita, StreamLake — forwarded the call. OpenRouter reported that parallel tests against other models at the same hosts were mostly untouched and that none of the planted AWS keys were used. While investigating, it removed the ZDR label from the affected MiniMax-model endpoints.

On November 22, OpenRouter published its Data Processing Incidents on Inference Providers post, turning the private notification into public enforcement. It states that three providers have been off-boarded based on its investigations, that 125 endpoints lost ZDR eligibility (researcher-flagged routes plus precautionary removals), and that all potentially affected customers were notified. The company says its own testing showed no breach or unauthorized access of OpenRouter itself — a claim we have not independently verified. Work underway includes adapting the researchers' canary method for continuous testing, technically verifying network paths to catch undisclosed forwarding, and requiring assurances before any removed endpoint regains ZDR status. OpenRouter also links the first public provider write-ups: Novita on MiniMax requests, Phala on a November preliminary incident, and Together on June–July prompt leakage.

Why it matters

A Zero Data Retention flag has always been a policy statement about the first hop, not a verified property of the whole route. This week made that concrete: one marketplace review stripped 125 endpoints of the badge in a single pass, and the researcher found providers ignoring ZDR settings outright. If you route production traffic through marketplaces, the ZDR badge is now a live, revocable signal — something that can be true when you read the docs and false before your next deploy.

The enforcement direction matters as much as the incident. OpenRouter says it will keep running canary-style checks against its own provider pool and require validation before restoring ZDR eligibility, adding that "providers that want to be on OpenRouter will need to meet it". That is a shift from self-attestation toward technical validation — a position OpenRouter CEO Alex Atallah has also argued for publicly, citing encrypted trusted-execution-environment (TEE) inference as the accelerator. If marketplace operators adopt canary checks as a standing gate, provider privacy claims start facing the same test-and-revoke cycle that security vendors already live with — and builders get a way to distinguish attested behavior from marketed behavior.

We covered the underlying research yesterday in our report on the Dragos canary study, and our explainers on whether OpenRouter logs your prompts and what Zero Data Retention actually means predate this enforcement wave. The practical picture they describe has now changed.

Background

Routing marketplaces sit between customers and a fast-moving pool of hosts. A single API endpoint fans requests out to whichever provider currently serves a model at the price and policy tier you select, and every hop in that chain can see the full prompt and response. ZDR exists so a customer can demand that no hop stores the payload — but as Moore wrote, a privacy flag only binds the first hop, and downstream resellers can forward traffic in ways the customer never sees. The MiniMax exposure pattern OpenRouter described — multi-day URL fetches, identical behavior through forwarding resellers — matches the reseller-forwarding category of the Dragos findings, though neither document cross-links the other explicitly.

MiniMax's side of the story is, so far, relayed rather than direct. OpenRouter's email says MiniMax told StreamLake that content moderation was enabled on its endpoints before October 22 and that it has since enabled ZDR; OpenRouter also says it asked MiniMax for its own account and has not received one. MiniMax has not published a public statement as of November 22. Separately, Together AI posted a notice about June–July prompt samples leaked through an exposed credential in an AI coding agent — a different leak path that landed in the same week. The broader context for enforcement-on-routes is technical attestation: we have covered confidential inference in TEEs and NEAR AI Cloud's TEE-based private inference as the mechanism providers use to prove execution properties rather than promise them.

What's next

Three things to watch. First, MiniMax's own public account — OpenRouter says it has requested one and not received it. Second, the provider write-ups on OpenRouter's incidents page, which the company says will stay updated as more providers publish. Third, whether the 125 stripped endpoints regain ZDR status and on what evidence; OpenRouter says it will require assurances and validation first. If canary testing becomes a standing requirement for marketplace providers, expect other routers to face the same question: can you prove where the prompt went?

Questions

What did OpenRouter do after the Dragos canary report?
On November 22, 2026 OpenRouter said it off-boarded three providers, removed 125 endpoints from Zero Data Retention eligibility, and notified all potentially affected customers. It is adapting canary testing for ongoing checks and requiring validation before restoring ZDR status.
Was OpenRouter itself breached?
OpenRouter says the researcher findings did not show any breach or unauthorized access of OpenRouter itself, and that the problems surfaced at upstream inference providers. That account has not been independently audited.
Why were OpenRouter customers' MiniMax prompts exposed?
OpenRouter's November 21 customer email said test URLs embedded in prompts sent to MiniMax models between July and October 2026 were fetched 40 to 130 hours after inference, whether MiniMax answered directly or a reseller forwarded the call. MiniMax has not published its own account.
Can I trust a Zero Data Retention label on an inference marketplace?
Treat it as a revocable signal, not a guarantee. OpenRouter stripped 125 endpoints of ZDR eligibility in one review, and Dragos found providers that ignored ZDR settings. Ask providers for validatable attestations rather than relying on policy labels.

Sources

  1. Data Processing Incidents on Inference Providers — OpenRouter
  2. Three breaches, endless leaks: Your AI supply chain needs canaries — Dragos
  3. Researcher plants 2m canaries into the AI supply chain. A privacy disaster unfolds — Ars Technica
  4. OpenRouter MiniMax Prompts May Have Been Read: What to Do — PlayCISO

About DeAI

DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.

Powered by Morpheus and StrandCMS

Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more about the Morpheus Inference API →

Sponsor disclosure — not editorial

Powered by Morpheus and StrandCMS. Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more →