OpenRouter, a service that routes AI requests to models run by many different hosts, has removed three of those hosts and told affected customers that their prompts may have been exposed. It also stripped 125 endpoints of Zero Data Retention eligibility, the setting that is meant to stop prompts and replies from being stored.
The OpenRouter platform itself was not breached, the company says. The trouble sits with the providers behind it, and with a privacy flag that, according to the research, does not always travel with the request.
OpenRouter’s incident notice follows a report from the security firm Dragos. Over five months from July through November 2026, Dragos submitted about two million prompts that carried unique links and fake credentials, then watched where they were used. Its activity logs pointed to three AI providers whose data had leaked, and none of them was aware until Dragos got in touch.
The biggest case involved a couple of hundred AWS keys planted in prompts. Every request sent during a few windows in late October was stolen, including some the provider rejected before any model ran, and the attacker then made tens of thousands of calls to AWS with them. In two of the breaches, some of the stolen requests had been marked as Zero Data Retention.
Dragos also found sellers on an inference marketplace who could read and keep prompts using ordinary logging, with no exploit needed. In another case, a link from a prompt flagged for Zero Data Retention turned up in a chat channel in China 24 days later.
Together AI is one of the providers that has disclosed an incident: an AI coding agent committed a credential to a public repository, and an unauthorized party used it to download stored prompt samples from June and July. Five of the credentials that party tested were Dragos canaries. OpenRouter’s notice also lists disclosures from Novita and Phala, and says not every affected provider has posted publicly.
OpenRouter says the 125 removals include precautionary ones, so they are not 125 confirmed breaches. It is checking network paths for requests forwarded to third parties and will want assurances before restoring any endpoint. If you received a notice, treat any password or API key you sent through the affected provider as exposed and rotate it. Dragos advises keeping secrets out of prompts altogether, since a privacy setting is only honoured by the first hop in the chain.



