VYPR
advisoryPublished Nov 22, 2026· 1 source

OpenRouter Users' Private AI Chats Potentially Exposed via Compromised Third-Party Providers

OpenRouter has revoked Zero Data Retention eligibility for 125 endpoints and removed three AI hosting providers after researchers found evidence of compromises and policy violations.

OpenRouter, a platform providing a unified API for accessing various AI models, has taken action against three of its AI hosting providers after researchers discovered evidence of compromised services and policy violations. The company revoked Zero Data Retention (ZDR) eligibility for 125 endpoints and removed the offending providers, citing concerns that user prompts and data may have been improperly stored or forwarded. This incident highlights the complex security challenges inherent in AI service aggregation, where data can pass through multiple third-party hands.

Researchers from Dragos conducted an extensive investigation between July and November 2026, sending nearly two million test prompts. These prompts contained unique links and canary tokens, allowing researchers to trace the data's path and detect unauthorized access or storage. The findings revealed that some providers copied test requests even when the AI model rejected them before processing. More critically, evidence suggested that some requests designated for Zero Data Retention—a policy intended to ensure prompts and replies are not stored—were instead retained or forwarded to other services.

One particularly concerning incident involved a test link sent to a U.S.-based provider that later surfaced in a company chat in China, approximately 24 days after its initial transmission. This suggests a potential for sensitive user data, including conversational content and potentially credentials inadvertently included in prompts, to be exposed to unauthorized parties.

OpenRouter has identified Novita, Phala, and Together as the providers in question. The company emphasized that the 125 ZDR revocations are precautionary measures and do not necessarily indicate confirmed breaches at each endpoint. OpenRouter is currently verifying whether these providers forward requests or replies to other companies and plans to reinstate ZDR eligibility only after thorough validation.

Customers who received a notification from OpenRouter are advised to review the content of their past prompts sent through the affected providers. If any prompts contained sensitive information such as passwords or API keys, users should immediately revoke or rotate those credentials and monitor their access logs for any suspicious activity. This incident echoes previous warnings about the importance of scrutinizing AI service providers' data handling practices.

To mitigate future risks, OpenRouter recommends that customers restrict sensitive workloads to only the most trusted and vetted providers. Furthermore, users should avoid including confidential secrets directly within prompts, even when relying on ZDR policies. The incident underscores the need for robust due diligence and continuous monitoring of third-party AI service providers to safeguard user privacy and data security in the rapidly evolving AI landscape.

Synthesized by Vypr AI
OpenRouter Users' Private AI Chats Potentially Exposed via Compromised Third-Party Providers · VYPR