A Chinese-speaking attacker has used autonomous AI agents to break into more than 100 websites, marking one of the clearest signs yet that offensive cyber work is moving beyond simple AI-assisted scripting. The campaign relied on agents to search for weaknesses, test them and continue attacking targets with only limited human direction.
The significance is not that the attacker invented a radically new technique. It is that work which used to consume hours of human attention could be delegated and repeated across many more targets at once. That lowers the cost of scanning for neglected systems and makes slow patching more dangerous.
For defenders, that means the time available between exposure and exploitation may keep shrinking. Internet-facing applications, APIs and remote-access systems that sit forgotten for weeks or months are increasingly likely to be found by automation rather than by a patient human operator.
The same risk is starting to appear inside organisations adopting their own AI agents. Systems that can browse files, call tools or act on behalf of staff need strict limits around credentials and permissions. An automated system should not be able to turn one mistake into a chain of privileged actions.
The practical lesson is straightforward: AI may change the speed of attack, but it does not remove the value of disciplined engineering. Accurate asset inventories, rapid patching, strong authentication, constrained privileges and reliable monitoring are becoming more important, not less.