OpenRouter Faces AI Chat Leak Over Provider Missteps

OpenRouter has come under scrutiny after internal research revealed potential exposures of private AI chats due to lapses by some of its hosting providers. While the company itself was not breached, evidence suggests that third-party endpoints may have mishandled requests—including those marked for zero data retention. That means private prompts and chats could have been stored or forwarded beyond what users expected. Recent actions now aim to limit further risk and restore user trust.

How the Exposure Happened

OpenRouter operates as an API gateway, routing users’ prompts to different AI model providers. Users assume data sent to a provider stays there, particularly when endpoints are flagged for zero data retention. But independent research involving test prompts—some embedded with unique tracking links or so-called “canary tokens”—found that three hosting providers failed to uphold these assurances. Not only were rejected prompts occasionally copied, but in at least one instance, data intended to be transient was resurfaced 24 days later in a chat within an entirely different organization based in China, despite originating from a U.S. provider.

OpenRouter’s Response & What You Should Do

Following the revelations, OpenRouter removed three providers—Novita, Phala, and Together—from certain service roles and revoked zero data retention status from 125 endpoints as a precaution. The company insists that these 125 are not confirmed breaches, but endpoint access has been frozen pending a more thorough verification process. Customers affected have been alerted directly.

If you’ve used any of the impacted providers, take steps to protect your data: review prompt history, revoke and rotate API keys or passwords if they were included, and monitor for any suspicious activity in associated logs. For tasks involving sensitive or confidential content, it may be safer to stick only with providers you’ve vetted, and never rely solely on data-retention labels.

Wider Implications for AI Privacy

This incident underscores long-standing concerns about AI model providers, especially those offering “zero data retention” promises. Transparency only goes so far when routing chains are opaque, and each added hand in the chain increases risk. As AI tools proliferate, organizations and users alike will need stronger controls—not only technical measures around data policies, but also processes to audit and enforce them.

What this teaches us is simple: guarantees of privacy and confidentiality are only as strong as the weakest link. OpenRouter’s swift action to remove flawed endpoints and notify users is a step forward. But real progress means building systems where “zero retention” isn’t just a label, but an enforceable reality. Companies offering AI APIs must adopt more rigorous audits, end-to-end transparency, and stronger accountability. Users in turn should remain vigilant—treat every service interaction as potentially exposed, and assume data could travel further than promised.