OpenRouter has recently taken significant measures to protect user privacy by cutting ties with three AI hosting providers and revoking Zero Data Retention eligibility from 125 endpoints. This action follows an investigation that uncovered potential data policy breaches and compromises among some third-party providers.
Details of the Security Incident
In a statement released on November 22, OpenRouter informed its users about the potential exposure risks. Despite the findings, there was no evidence of a direct breach within OpenRouter’s infrastructure itself. The company has not specified the number of users or conversations that might have been affected.
OpenRouter provides its customers with a unified API to access various AI models hosted by different providers. However, the initial recipient of a user request may not be the sole handler, as requests could be passed along to other services, potentially exposing sensitive information.
Compromised Providers and Data Retention Issues
According to a report by cybersecurity firm Dragos, nearly two million test prompts were sent between July and November 2026. These tests, embedded with unique identifiers known as canary tokens, revealed security lapses. Instances of compromised data were traced back to three hosting providers, with unauthorized copies of requests noted even before processing.
Moreover, some requests labeled under the Zero Data Retention policy, which mandates non-storage of prompts and replies, were improperly retained or shared. For example, a test link directed to a U.S. provider was discovered in a Chinese company chat 24 days later, highlighting the potential reach of these breaches.
User Precautions and Future Outlook
OpenRouter’s disclosures identified Novita, Phala, and Together as implicated providers. While the removal of endpoints from Zero Data Retention eligibility was precautionary, it underscores the necessity for ongoing vigilance. OpenRouter is actively verifying endpoints before reinstating their eligibility status.
Users who received notifications should assess their data submissions through affected providers. In cases where sensitive information such as passwords or API keys were included, users are advised to revoke or rotate these credentials and scrutinize their access logs for unusual activity.
As demonstrated in previous cybersecurity reports, reviewing both chat histories and credentials is crucial. To mitigate future risks, users should limit sensitive operations to trusted providers and avoid embedding secrets in prompts.
Stay informed about how leading organizations are addressing AI and cybersecurity challenges and learn from their strategies. View details to safeguard your data and operations.
