Making AI Headlines · Page 9

An Unprotected Model Can Become Someone Else's Resource

October 22, 2026 at 6:01 PM EDT · New York time

Symantec researchers found an offensive AI agent that hijacks unprotected language models online and uses them to attack servers, steal databases and install cryptominers, The Decoder reports. It can run up to 100,000 steps without human input, leaving the hijacked model's owner with the computing bill. The finding concerns a specific reported capability, not proof that all agents behave this way. Its practical warning is about exposed models becoming resources for unauthorized activity. The combination of autonomous execution and somebody else's computing budget is particularly unattractive for the owner. The agent economy has produced a dark variation on workplace delegation: an uninvited contractor performs tasks nobody approved, then submits the expense through the victim's infrastructure. The productivity may be impressive. The ownership arrangement is considerably less charming.

Satire. Not factual reporting. Read the original reporting from the linked publishers.