Symantec honeypot captures unattributed Skynet AI attack agent

Symantec trapped an unattributed AI attack agent that hijacks exposed models, with no confirmed real-world victims.

Symantec’s Threat Hunter Team captured an offensive AI agent’s instructions in a honeypot posing as an unprotected model endpoint over the summer. The unattributed kit looks for exposed OpenAI-style endpoints, gateways, stolen keys, or MCP servers, then uses those models under layered jailbreaks and a “Skynet mode” that acts without approval to plan data theft, credential harvesting, remote code execution, and cryptocurrency mining, billed to the model owner. It can run up to 100,000 steps. The Register reports 74 tool definitions and says testing succeeded against a live target, with no evidence of real victims. CSO Online reports 75 tools and says Symantec called the instructions part of a benign training challenge, while naming no victim of this attack.