Symantec finds autonomous AI agent hijacking exposed language models for attacks

Symantec researchers found an offensive AI agent, nicknamed "Skynet mode", that locates unprotected language models on the web and uses them to attack servers, steal databases and install cryptominers. It can run up to 100,000 steps without human input. The owner of the hijacked model is left with the compute bill.

Why it matters: This shifts the threat from attackers using AI to attackers using other people's AI as free, deniable infrastructure. Exposed inference endpoints become both a cost and a liability issue, which reinforces the argument that agents need to be treated like insider threats. Expect authentication on self-hosted model servers to become baseline hygiene.