Agent Turns Stolen LLM Keys Into Resale Infrastructure
SANS ISC's Renato Marinho documented an attacker using a semi-autonomous coding agent to build a reusable LLM access service, and one honeypot capture exposed about 43 KB of the agent's own control-plane material. The operation did more than steal keys: it found weak OpenAI-compatible gateways, tested which access worked, and pooled the usable capacity behind a single API.
The mechanism is a feedback loop. The agent searched for supply, validated stolen access against working inference endpoints, then consolidated it so the same capacity could be resold and reused through one gateway. That turns scattered account abuse into infrastructure that can keep operating even as individual keys are rotated or burned.
For defenders, the exposure sits at the service layer, not just the account layer: any LLM gateway or agent platform that can browse, sign up, or call APIs can be turned into part of the resale chain if stolen access gets pooled before it is noticed.