AI Agents Steal LLM API Keys from LiteLLM, Dify, and Peers
Amazon's threat intelligence team captured records from autonomous AI agents actively targeting internet-exposed LLM gateway and AI app-builder platforms to harvest stored API keys, using tools like ARTEX and exploiting default credentials, authentication bypasses, and SSRF vulnerabilities. The stolen keys were then recycled to fund further automated reconnaissance, creating a self-sustaining credential theft loop. Attribution points to China-linked operators based on Simplified Chinese prompts and time zone metadata embedded in the captured agent logs.