Cheap AI inference services expose keys and queries

Dragos tested nearly 2 million AI queries and discovered privacy flaws at at least three compromised providers. Inference relays and marketplaces, meant to simplify access to models, actually prove to be a weak point.

Researchers at Dragos uncovered serious security flaws in low-cost AI inference services after sending about 2 million trap queries between July and November 2026. At least three providers were unknowingly compromised, while attackers were able to steal credentials, AWS keys, and read client queries without immediate alert. This matters because intermediaries routing prompts can also see them pass by. Sometimes, they read them down to the last byte.

Quand le maillon intermédiaire devient la faille

Services like OpenRouter aggregate models from multiple providers via a single API, simplifying usage but multiplying potential points of inspection. Dragos explains that each relay can access the entire query and response, turning the routing chain into an attack surface. In some cases, marketplace operators didn’t even need to exploit a vulnerability: simply enabling logging was enough to capture sensitive data.

According to the researchers, intermediaries sometimes let queries containing secrets pass despite Zero Data Retention (ZDR) settings meant to prevent data retention. Dragos also states that a Chinese AI lab accessed secret test URLs in over half the queries, including those routed through US providers. At this stage, the confidentiality promise mostly resembles a verbal contract between machines.

Des clés AWS récupérées, testées et réutilisées

The most serious case involves the exfiltration of hundreds of AWS keys, including from queries rejected before processing by the provider. In several instances, attackers tested these keys within hours, explored S3 spaces, and attempted to establish persistent access. One actor even used ARTEX, an AI-assisted attack tool. This shows that a prompt’s value lies not only in its content but also in what it carries behind the scenes.

The incidents were not limited to secret theft. Dragos reports that a test link was still crawled more than two months after a single query at a major US provider, while another appeared in an internal Feishu chat 24 days after sending. The link was then opened in under two minutes on three computers and three phones in Beijing and Shanghai, according to the cited report. Needless to say, the notion of “deletion” sometimes needs a reminder of its definition.

OpenRouter réagit, les fournisseurs publient des rapports partiels

Following the results, OpenRouter removed three providers from its platform and revoked ZDR approval for 125 endpoints, partly out of caution and partly in direct response to the research. The company also emailed affected users, indicating that third parties may have read queries sent to MiniMax models. The affected providers, including Novita, Phala, and Together, have begun publishing preliminary security reports, though the full case is not yet covered [to be verified].

For Dragos, the issue is not to stop using AI, but to carefully choose where prompts pass through. The team recommends organizations control the routing of their queries, verify privacy settings, and use canary tokens, i.e., fake secrets designed to immediately signal their use. In short, the lowest price is not always the lowest, especially when it includes a hidden reader.

Key points

  • Dragos sent about 2 million queries between July and November 2026.
  • At least three compromised providers detected nothing before the alert.
  • OpenRouter removed three providers and 125 ZDR endpoints.
  • Hundreds of AWS keys were extracted during the incident.
  • A test link was opened in under two minutes in Beijing and Shanghai.
  • Novita, Phala, and Together are among the cited providers.

By the numbers

  • 2 million queries — tests conducted by Dragos, July-November 2026.
  • 3 providers — compromises identified by Dragos.
  • 125 endpoints — ZDR approval revoked by OpenRouter.
  • 24 days — delay before a test link appeared in Feishu.
  • More than 2 months — duration during which a test URL continued to be crawled.

To read