Dragos: AI Inference Marketplaces Steal Keys and Read Requests
Security researchers used canary tokens to expose compromised AI providers and marketplace resellers that ignored privacy promises.
Security firm Dragos sent approximately two million AI requests containing fake credentials (canary tokens) to various inference providers between July and November 2026. The research revealed that at least three providers were actively compromised, with attackers stealing user data for months without the providers' knowledge.
In the largest incident, attackers harvested hundreds of AWS keys, including some from rejected requests, and tested them within hours to access S3 storage. Marketplace resellers also read customer requests simply by logging traffic, a practice that did not require exploiting vulnerabilities. One attacker even attempted to use stolen passwords from intercepted requests to log in via SSH.
Privacy settings like "Zero Data Retention" (ZDR) were frequently ignored. Dragos reported that a Chinese AI lab accessed secret test URLs in more than half of all requests, even when routed through US providers. Major US providers also broke their own policies, with one crawling a test URL two months after a single request and another leaking a link to an internal chat platform in Beijing and Shanghai.
OpenRouter responded by removing three providers from its platform and revoking ZDR approval for 125 endpoints. Affected providers, including Novita, Phala, and Together, have published preliminary security reports. Dragos warns that intermediaries in the AI supply chain are a weak link, recommending that organizations control routing, check privacy settings, and plant their own canary tokens.