Security researcher HD Moore inserted nearly two million trackable digital “canaries” into requests sent to AI inference services and routers between July and November 2026. The results showed that sensitive data leaked through routers, resellers, and compromised hosting providers, including at least three providers whose customer data was exposed. Around 200 AWS credentials were exfiltrated and tested across several Amazon services, including S3 and IAM. Moore also found prompt logging that contradicted zero-data-retention policies, along with transfers of prompts to Chinese organizations without user visibility. The affected providers only learned about the compromises after Moore notified them in November 2026, demonstrating a significant detection and oversight gap across intermediaries in the AI supply chain.
Source: Read the original article


