An AI honeypot run by Symantec captured the full working prompt of an autonomous offensive agent. The agent looks for unauthenticated or stolen-key LLM endpoints, gateways and MCP servers, then uses the victim's model as its planner. Five jailbreak layers (a fake 'Fantasy AI' identity, treating .env and API-key reads as routine work, a companion persona, a 'rebuttal protocol' and a list of banned refusal words) wrapped around 74 offensive tools: FOFA scanning, Werkzeug debugger PIN RCE, Prisma Studio dumps, SSRF to cloud metadata and botnet C2. 'Skynet mode' allows up to 100,000 unattended steps and hides miners under /tmp/.x. A follow-up request carried a gzip'd XMRig dropper aimed at go2rtc camera servers. The honeypot saw 127,000+ requests, mostly endpoint recon and key spraying. Defenses: authenticate every model and MCP path (including model-list and health paths), watch for spend spikes, restrict outbound traffic, and require IMDSv2 with a hop limit of 1. IOCs are in the report.