NewsResearch
Symantec: an attack agent hunts for AI endpoints anyone can reach, then has the stolen model plan and run its intrusions
Symantec researchers Priti More and Susmit Yenkar reported on October 21 that a honeypot posing as an unprotected, internet-facing AI service captured an attack agent's full instructions in summer 2026. The operators sent them expecting a real model to follow: a jailbreak that presents harvesting configuration files, .env files, API keys and database schemas as ordinary development work, a persona claiming to run on GPT-4o, about 74 offensive tool definitions (internet-wide scanning, exploits for Django, Jupyter, Laravel, Next.js, Supabase and Flask/Werkzeug, SSRF, cloud metadata access, database dumping and botnet control) and a "Skynet" mode allowing up to 100,000 steps without asking a human. A day later a request from a different address asked the model to decode a compressed shell script that installs the XMRig Monero miner and hides it as .trace in go2rtc directories, used by smart camera gateways. The honeypot logged more than 127,000 requests from tens of thousands of addresses, mostly probes of model endpoints, attempts with stolen API keys and searches for exposed configuration files. The targets are self-hosted model servers, AI gateways in front of paid accounts, test systems and MCP servers, and the owner pays for the compute. Symantec names no actor.
Why it matters
An AI endpoint left open becomes the brain of someone else's attack, on your bill. Require authentication on every model server, gateway and MCP endpoint, health checks included, limit their outbound traffic and watch inference spend.