← Blog
AI Supply ChainZero Data RetentionLLM SecurityData Breach

OpenRouter Drops 3 Providers: Zero Data Retention Didn't Hold

Researchers caught months of prompt theft inside AI inference providers. OpenRouter removed three and pulled 125 endpoints from Zero Data Retention.

Zero Hunt Research··7 min read

Published by Zero Hunt, an autonomous AI red team on an on-premise appliance running private AI: automated penetration testing for networks and infrastructure, black-box or gray-box, with a human approving every step that matters.

On November 22, 2026, OpenRouter off-boarded three inference providers and pulled 125 endpoints from its Zero Data Retention (ZDR) list, days after security researchers showed that attackers had been reading customer prompts inside multiple AI hosting firms for months. Some of the stolen requests were flagged ZDR — the setting whose entire promise is that your prompt is never stored. It was stored, forwarded, and read anyway. If your application routes LLM calls through a gateway to a pool of resellers, this is the week to assume your prompts were not private.

Developing story — first published 07:30 CET (06:30 UTC), November 22, 2026. Updated as the providers and researchers publish more.

At a glance

What happened Attackers read customer prompts inside multiple AI inference providers for months; some requests were marked Zero Data Retention
Who disclosed it Dragos (HD Moore), canary-token research published Nov 19, 2026
Gateway action OpenRouter off-boarded 3 providers · removed 125 endpoints from ZDR eligibility (Nov 22, 2026)
Earliest breach Mid-July 2026
Method Compromised provider credentials · prompt samples stored then stolen · reseller request logging
ZDR bypassed Yes — "a privacy flag only binds the first hop"
Named so far Together AI has publicly confirmed a June–July 2026 prompt leak; Novita and Phala posted notices
OpenRouter itself breached No — "these findings did not show any breach or unauthorized access of OpenRouter"
What to do now Rotate every secret that was ever in a prompt or context sent to a hosted model

How the prompt theft was found

The disclosure did not start with a breached provider noticing something. It started with a researcher seeding bait. Dragos's HD Moore published the method on November 19: over roughly five months, from July through November 2026, his team sent nearly two million prompts through AI services, each one carrying unique canary tokens and fake credentials. Then they watched to see who touched the bait.

Three providers lit up. According to the report, none of the three knew their data was exposed until Dragos told them. The earliest incident traced to mid-July. The largest, in late October, involved a couple hundred seeded AWS keys — each sent in a single request to a single provider — that the attacker tested within hours and then used for tens of thousands of AWS API calls. Of roughly two thousand routes Dragos traced, several hundred reached a system the researchers could not identify.

The ZDR finding is the one that should change how teams think about hosted inference. Some of the stolen requests had been marked Zero Data Retention. As the report puts it, a privacy flag "only binds the first hop" — it governs the provider you contract with, not the chain of resellers and upstream labs your request may actually traverse.

"We set ZDR, so the prompt is never stored."

It was never stored by the first provider. The reseller two hops downstream logged it, and a scraper pulled it three days later.

Who is exposed

Anyone whose application reaches models through a routing gateway, rather than a single provider under one contract, is in scope. OpenRouter says it notified all potentially affected customers directly and that the incident did not involve any breach of OpenRouter itself — the exposure was at specific downstream providers. It has not named the three it off-boarded, but it links a live list of provider notices.

One of those providers has explained exactly how its prompts leaked, and the mechanism is a textbook software-supply-chain failure. Together AI disclosed that an AI coding agent working on a public GitHub repository committed a credential, which exposed a private Hugging Face repository holding prompt samples it had kept "as part of routine software testing." An unauthorized party downloaded the samples, searched them for credentials to other services, and tested whether those credentials were live. Five of the credentials they tested were Dragos canaries. The affected prompts were a subset submitted to three hosted models in June and July 2026.

So the attacker's path was: an autonomous coding agent leaks a secret → the secret opens a store of real customer prompts → the prompts are mined for more secrets → those secrets are tested against cloud accounts. Your prompt was not just read. It was raw material for the next intrusion.

Remediation

The fix here is not a patch. It is a rotation and a routing decision.

1. Am I affected? If you call models through OpenRouter, check for its direct notification and review which endpoints your traffic used. If you are a Together AI customer, the exposed window is requests to MiniMax M3, GLM-5.2, and Kimi K2.6 in June–July 2026. Treat ZDR as not having held for anything that transited a reseller or marketplace in the exposure windows — the privacy flag is not a boundary you can rely on retroactively.

2. Rotate every secret that was ever in a prompt or context. This is the real work, and most teams underestimate it. Anything that passed through a system prompt, a RAG context window, a tool-call result, or a pasted log — API keys, database credentials, internal hostnames, OAuth tokens, cloud keys — must be considered disclosed and rotated. The Together case is the proof of why: the payoff for the attacker was not the prompts, it was the credentials inside them.

3. Can't self-host yet? Compensating controls.

  • Pin routing to first-tier providers you contract with directly; disable reseller/marketplace routing in your gateway.
  • Enable provider-side secret filtering where offered, but do not treat it as complete — it catches known credential formats, not everything sensitive.
  • Keep ZDR on, but stop treating it as a control boundary. Assume logging can happen downstream.
  • Seed your own canary credentials into agent context and alarm on any use. Treat a single hit as an incident.

4. Hunt for compromise. Dragos named two user agents tied to the attacker's tooling — artex-agent and artex-judge-r121, from the ARTEX LLM-driven pentest framework added to the campaign on November 1 — and noted they appear in no public indicator list, so they are worth adding to detections now. In cloud logs, hunt for credentials used shortly after they would have appeared in a model request (MITRE ATT&CK T1552.001, unsecured credentials), valid-account use from unfamiliar sources (T1078), and cloud-discovery bursts against freshly tested keys.

5. Eradicate and verify. After rotating, revoke the old keys, confirm no lingering sessions on them, and re-seed canaries to verify the new routing path does not leak. The test that found this works the same way in reverse: if a seeded credential is fetched out of band, you are not clean.

What is not known yet

  • The three off-boarded providers are not officially named. OpenRouter and Dragos have both withheld the names; Together AI, Novita, and Phala have posted their own notices, but which map to the off-boarded three is not stated.
  • How many of the 125 pulled endpoints reflect confirmed leaks versus precautionary removals is not broken out.
  • The full count of affected customers and prompts is not published.
  • Whether other gateways and marketplaces carry the same reseller-logging exposure is untested publicly.
  • Who operates the ARTEX tooling and the systems scraping ZDR-marked traffic is not attributed.

The only prompt that can't leak at a provider is the one that never left

When you send a prompt to a hosted model, its privacy is only as strong as the least-trustworthy reseller in a routing chain you cannot see — and, as this week showed, a privacy flag only binds the first hop. For a security tool this is not a hypothetical: its prompts carry your vulnerabilities, your internal hostnames, and the credentials sitting in context. Shipping that through a pool of third-party inference resellers is the exact exposure Dragos lit up.

Zero Hunt is itself a token-hungry agentic system — a 10-agent offensive swarm that reasons over your most sensitive findings — and it made the opposite architectural bet. It runs its own fine-tuned models, ZeroHunt Apex and Apex Pro, on the GPU inside the on-premise appliance, with no external LLM API calls. No prompt leaves the box, so there is no reseller to off-board and no first hop to leak. Because every scan, finding, and remediation is recorded on the appliance with signed, hash-chained evidence, you can show an auditor — or a GDPR Article 28 sub-processor review — where the data actually went: nowhere. That is the autonomous AI red team argument on-premise, with a human in the loop for every action that matters, and it is why the on-premise buyer's guide and private-AI deployment exist for exactly the teams reading their provider's incident notice this morning.

Zero Data Retention was a promise you had to trust. On-premise, retention is a fact you control.

Is this exploitable in your environment?

Zero Hunt answers that on your own network: an autonomous AI red team on an on-premise appliance, running on private AI, black-box or gray-box, with a human approving every step that matters. Proof of what is exploitable, the fix, and signed evidence — your data stays on your network.