OpenRouter, the gateway that routes millions of AI requests to outside model providers, has cut three providers from its platform and stripped 125 endpoints of their "zero data retention" status after researchers showed private prompts were being copied, stored and read by people never meant to see them.

The company announced the action on Sunday, November 22, in an incident notice titled "Data Processing Incidents on Inference Providers." It said it had notified every customer who may have been affected, and stressed that the findings showed no breach of OpenRouter itself — the problem was downstream, at the providers who actually run the models.

Two million planted secrets

The findings come from industrial cybersecurity firm Dragos, which published its research on November 19. From July to November, Dragos seeded about two million prompts with unique web links and credentials — canary tokens designed to reveal themselves if anyone copied, stored or used them — then watched what happened.

At least three AI hosting providers showed evidence of compromise or of failing to honour their own data policies. The earliest breach began in mid-July, Dragos said, and none of the three knew they had been hit until the researchers told them.

One unnamed Chinese AI lab fetched the secret links from more than half of the seeded prompts sent to its models between July and October — including about two-thirds of the requests marked for no data collection or zero data retention. In one case, a secret link from a planted prompt turned up in a Feishu chat 24 days after it was sent — and was opened within two minutes from three desktops and three phones in Beijing and Shanghai.

The planted prompts were bait, Dragos said, but they ran alongside real customer requests — almost certainly copied too.

Three providers named

OpenRouter's notice links security announcements from three providers. Novita said a routing error sent some MiniMax requests to MiniMax's official API instead of Novita-hosted infrastructure between May 28 and November 11. Phala said an attacker took control of one of its inference servers and could view request data while it was being processed; DeepSeek V4 Flash requests routed through OpenRouter between October 21 and November 8 should be treated as exposed, and any secrets in them rotated. Phala contained the incident on November 8. Together AI disclosed that an unauthorised party downloaded stored prompt samples after one of its coding agents committed a credential for a private repository to a public GitHub repository in June and July.

The 125 endpoints removed from zero data retention eligibility include providers flagged by the researchers plus others removed as a precaution, OpenRouter said — the removals do not mean all 125 were confirmed breached. The company did not say how many users or conversations may have been exposed.

What happens next

OpenRouter said it is adapting the researchers' canary-token testing so the checks can run on an ongoing basis, and is verifying network paths to find where requests and responses may be forwarded to third parties. Endpoints stripped of zero data retention status will only get it back after what the company called appropriate assurances and validation.

"We take data policy enforcement very seriously," the company said, adding that providers that want to stay on OpenRouter "will need to meet" a higher bar.

For anyone who routes sensitive work through an AI gateway, the lesson is blunt: a "zero data retention" promise is only as strong as the weakest provider behind it — and, for months, that weakness was being watched from the outside.