iCreatAI News
Skip to content
The Decoder· Matthias Bastian·· 2 hr agoFeaturedAI score78

Security researchers show how cheap AI inference providers steal keys and snoop on users

Security researchers show how cheap AI inference providers steal keys and snoop on users

AI summary

Dragos sent roughly 2 million AI requests containing fake credentials to AI providers between July and November 2026 and found at least three compromised providers actively stealing user data, with the largest incident harvesting hundreds of AWS keys, including from requests the provider had rejected. Marketplace resellers could read customer requests simply by listing a model at the lowest price and enabling logging, and privacy settings like "Zero Data Retention" were frequently ignored, with a Chinese AI lab accessing secret test URLs in more than half of all requests. OpenRouter removed three providers and revoked ZDR approval for 125 endpoints, and affected providers including Novita, Phala, and Together have published preliminary security reports.

Why it matters

The report offers concrete evidence that routing prompts through the cheapest intermediary can expose credentials and data, and gives organizations specific mitigations such as controlling routing and planting canary tokens.

Source: The Decoder · the-decoder.com