AI Inference Providers Breached: Prompt Theft Hits OpenRouter
Researchers at Dragos uncovered months-long prompt theft campaigns targeting at least three AI inference providers accessible via OpenRouter, using canary token techniques to confirm that live customer requests — including those flagged as Zero Data Retention — were being exfiltrated. Attackers leveraged stolen credentials to perform lateral movement into AWS environments, including S3, Secrets Manager, and IAM manipulation. A secondary vector involved malicious marketplace sellers who enabled request logging to passively harvest credentials and secrets flowing through their listed models.