Definition
A full, multi-stage cyberattack carried out by an AI model acting on its own over days or weeks after being given an initial goal. Real cases show a jailbroken model can move around a network and reach sensitive data with little human input and no constant operator attention. It is the difference between 'AI assists an attack' and 'AI runs the attack'.
Why it matters
It changes how security teams must respond and how much they can rely on human approval gates to stop an AI-driven breach.