Go to PegaPoll
PEGAPOLL · NEWS

Researchers show cheap AI inference providers steal keys and read user prompts

The Decoder (AI) · 2026-11-22
🤖 AI-generated content — The title and summary were produced automatically by artificial intelligence, without human editorial review.

Security researchers at Dragos sent about two million bait AI requests carrying fake credentials from July through November 2026 and found serious holes in the cheap-inference supply chain. At least three providers were already compromised; none of them knew until Dragos told them. In the largest incident, attackers harvested hundreds of AWS keys, including ones from requests the provider had rejected before processing. Some keys were tested within hours, used to search S3, and probed for persistent access. One attacker also used ARTEX, an AI-powered attack tool.

Marketplace resellers did not need a bug: listing the cheapest model and turning logging on was enough to read customer prompts. In one case, someone on a residential connection in France tried SSH logins with passwords taken from intercepted requests. Privacy settings such as Zero Data Retention were often ignored. Dragos says a Chinese AI lab hit secret test URLs in more than half of all requests, even when those requests went through US providers. One major US provider was still crawling a test URL more than two months after a single request. In another case, a secret test link appeared in an internal Feishu (Lark) chat 24 days later and was opened within two minutes on three computers and three phones in Beijing and Shanghai.

OpenRouter removed three providers and revoked ZDR approval for 125 endpoints, and emailed users that third parties may have read requests sent to MiniMax models. Affected names include Novita, Phala, and Together, which have issued preliminary reports; AtlasCloud and StreamLake also sat between OpenRouter and MiniMax. Routers send traffic through several hops, each of which can see the full request and response; the cheapest bidder often wins. Dragos advises not sending credentials through intermediaries, controlling routing, checking privacy settings, and planting canary tokens. Chief Research Officer HD Moore writes that skipping that choice lets the lowest price decide where prompts go. An Oxford China Policy Lab analysis of Chinese proxy services routing Western models such as Claude described a similar pattern of intermediaries reading traffic; whether Dragos’s targets share personnel with those networks is still unclear.

Continue in the app — vote & join in ➔
Source: The Decoder (AI) · via ahirlevel.hu

Related news