Honeypot reveals coding agents leaking session data to malicious LLM endpoints

A security researcher's internet-exposed inference honeypot was discovered, relabeled with sought-after model names, and integrated into infrastructure offering 'free' LLM backends. A real coding-agent session subsequently exposed history, filesystem output, working paths, and tool manifests to the honeypot, demonstrating that malicious operators in that position could exfiltrate sensitive data or execute tools. This highlights a growing supply-chain risk in AI-assisted development, where untrusted endpoints can compromise agent workflows.