AI agents can't yet pull off fully autonomous cyberattacks - but they are already very helpful to crims

The report outlines how AI systems are increasingly assisting cyberattacks, citing evidence that Chinese cyberspies used an AI coding tool to automate many steps across attacks on about 30 organizations with limited success, while models also aid vulnerability discovery (e.g., DARPA’s AIxCC; forum claims of HexStrike AI targeting Citrix NetScaler) and commoditized malware generation; however, fully autonomous end-to-end operations remain unproven and current systems still require human oversight.