Dragos reveals AI providers compromised, hundreds of AWS keys stolen in major breach
Security researchers at Dragos uncovered serious vulnerabilities in the AI supply chain, revealing that at least three AI providers had been compromised, resulting in hundreds of stolen AWS keys and unauthorized access to user data.
The investigation, conducted between July and November 2026, involved sending approximately two million fake AI requests to various providers. Dragos discovered that attackers had been actively stealing user data, with some providers unaware of their compromised status until notified by Dragos. In one significant breach, hundreds of AWS keys were harvested, including from requests that had been rejected prior to processing.
The researchers noted that attackers were quick to act, testing stolen keys within hours and attempting to establish persistent access to systems. One attacker utilized ARTEX, an AI-powered attack tool, while marketplace resellers exploited the system by simply reading customer requests. Dragos reported that these resellers did not need to breach security measures; instead, they could list models at the lowest price and enable logging to gain access to sensitive information.
In a troubling finding, a Chinese AI lab accessed secret test URLs in over half of the requests, even when routed through U.S. providers. Major U.S. providers also failed to adhere to their own privacy policies, with one continuing to crawl a test URL more than two months after it was initially requested. In another instance, a secret link appeared in an internal chat on Feishu, a Chinese collaboration tool, just 24 days after the request was made, and was opened within minutes on multiple devices in Beijing and Shanghai.
In response to these findings, OpenRouter removed three compromised providers from its platform and revoked Zero Data Retention approval for 125 endpoints. The company also alerted affected users that third parties may have accessed requests sent to MiniMax models. Affected providers, including Novita, Phala, and Together, have begun publishing preliminary security reports, although not all have completed their assessments.
Dragos advises organizations to be cautious when sending requests containing credentials to intermediaries and recommends that they control how their AI requests are routed. The researchers also suggest implementing canary tokens—fake credentials that alert users when they are accessed—to enhance security measures against potential breaches.
Why it matters: This breach raises significant concerns about the security of AI supply chains, particularly as companies increasingly rely on third-party providers for AI services. Organizations must prioritize security protocols and consider implementing measures like canary tokens to detect unauthorized access.
Original reporting: The Decoder