Skip to content

OpenRouter drops 3 providers after Dragos exposes prompt theft

Dragos seeded 2M canary prompts and caught AI inference providers leaking them for months — some marked Zero Data Retention. OpenRouter removed 3 providers and 125 ZDR endpoints.

Published 3 min read

Industrial-security firm Dragos says attackers stole data from inside AI inference providers for months, and none of the affected providers knew until researchers told them — with some of the leaked requests marked Zero Data Retention (ZDR). The primary is Dragos's November 19 report, "Three breaches, endless leaks: Your AI supply chain needs canaries," by chief research officer HD Moore. In response, the routing gateway OpenRouter removed three providers and 125 ZDR endpoints.

How Dragos caught it

Dragos seeded roughly two million prompts between July and November, each carrying a unique canary URL and canary credentials (starting with Thinkst Canarytokens, then a dedicated service). Any use of a canary meant someone had read the prompt that carried it. The findings:

  • At least three hosting providers had request data stolen; the largest theft took a couple hundred planted AWS keys in late October.
  • Multiple reseller marketplace sellers listed models at the lowest price, enabled request logging, and collected traffic — several appearing to run from home connections.
  • An unnamed Chinese AI lab scraped canary URLs from more than half of the prompts sent to its models between July and October, including about two-thirds of requests marked no-data-collection or ZDR.

Dragos reports that on November 1 an attacker used ARTEX, an LLM-driven penetration-testing framework, to confirm stolen keys were still live. The report flags two user-agent strings as detection candidates:

# User agents Dragos reports as absent from public indicator lists
artex-agent
artex-judge-r121

The three disclosed incidents

OpenRouter has not named the providers it removed, but three disclosed incidents of their own:

  • Phala: an attacker controlled an inference server and could view request data during processing; contained November 8. Requests to DeepSeek V4 Flash 0731 from Oct 21 00:00 UTC to Nov 8 05:35 UTC should be treated as exposed. (report)
  • Together AI: an AI coding agent committed a credential to a public GitHub repo; an unauthorized party downloaded stored prompt samples from June and July. Affected models included MiniMax M3, GLM-5.2 and Kimi K2.6; five tested credentials were research canaries. (notice)
  • Novita: a routing misconfiguration forwarded some MiniMax requests to MiniMax's official API, beyond what customers intended.

OpenRouter says the findings show no breach of OpenRouter itself, that it notified potentially affected customers, and that it now runs the researchers' checks continuously.

Action checklist

  1. If you route inference through OpenRouter or a reseller, treat any prompt or key sent in the exposure windows as compromised and rotate affected secrets.
  2. Stop relying on ZDR / "no retention" flags as a security boundary — Dragos found they bind only the first hop.
  3. Prefer first-tier providers directly over reseller marketplaces; verify where your requests actually go.
  4. Seed your own canary credentials (e.g., Thinkst Canarytokens) in prompts as tripwires.
  5. Add the ARTEX user agents above to detections.

Context

This is the AI-inference layer inheriting the same supply-chain exposure that has hit software registries and CI pipelines: trust delegated to an opaque chain of third parties, with no way for the customer to see the hops. Dragos ties the activity to the October intrusion at platform company Railway, which said its breach was likely part of a series targeting platform companies to reach AI inference providers. The privacy flag that reassured buyers — ZDR — turned out to describe a promise, not an enforced control.

Related stories