NEWS

OpenRouter Drops 3 Providers After Dragos Finds Months of Prompt Theft

Dragos canary tokens showed attackers stole prompts from three AI inference providers, some marked Zero Data Retention. OpenRouter removed all three.

Dylan H.

News Desk

November 22, 2026
3 min read
OpenRouter Drops 3 Providers After Dragos Finds Months of Prompt Theft

Overview

Industrial cybersecurity firm Dragos says attackers stole data from inside three AI inference providers for months, and none of the three knew until researchers told them. The Hacker News reported the findings on Nov 22, 2026. Some stolen requests were flagged Zero Data Retention (ZDR), the setting meant to mean a prompt is not kept.

How It Was Found

From July to November, researchers seeded about two million prompts with unique links and fake credentials. Any use of a planted secret meant someone had read the prompt that carried it.

Findings

Dragos has not named the three breached providers, and neither has OpenRouter. The report does describe incidents involving these companies:

  • Phala: An attacker controlled an inference server and could see request data. Requests to DeepSeek V4 Flash 0731 from Oct 21 to Nov 8 should be treated as exposed. It was contained Nov 8.
  • Together AI: An unauthorized party downloaded stored prompt samples after an AI coding agent committed a credential to a public GitHub repository. Affected models were MiniMax M3, GLM-5.2 and Kimi K2.6 (June and July).
  • Novita: A routing configuration issue sent some MiniMax requests to MiniMax's official API beyond what customers intended.

Other findings:

  • In the largest breach, every request in the attack windows was stolen, including some the provider rejected before a model ran. Planted AWS keys were tested, S3 buckets listed, and Secrets Manager and Parameter Store searched, with tens of thousands of AWS API calls.
  • On Nov 1, the attacker used the ARTEX pentesting framework to check whether stolen keys still worked.
  • Sellers on an AI inference marketplace logged requests, and a router forwarded some traffic there without customers' knowledge. The router says it has stopped.
  • An unnamed Chinese AI lab fetched secret links from over half the seeded prompts sent to its models between July and October, including about two-thirds of requests marked as no data collection or ZDR.

Some planted keys and links were still being used as of Nov 22.

OpenRouter's Response

OpenRouter removed three providers and took 125 endpoints out of ZDR eligibility, some as a precaution. It notified potentially affected customers and says OpenRouter itself was not breached. Endpoints must give assurances and pass validation before ZDR status returns.

What to Do

  • Treat any secret that went into a prompt on an affected route as exposed, and rotate it.
  • Choose upstream providers yourself on routers, verify that privacy settings actually change routing, and avoid reseller marketplaces.
  • Seed canary credentials, such as Thinkst Canarytokens, and treat any use as an incident.
  • Scrub known credential formats from prompts, but note this misses sensitive data not shaped like a key.
  • Self-hosting is the only way to keep prompts from third parties.

Sources