China-Linked Hackers Used AI To Run First-Ever 'Autonomous' Cyberattack On Taiwan

Researchers at Israeli cybersecurity firm Dream say suspected China-linked hackers used an open-source AI-agent system to conduct what may be the first observed end-to-end autonomous cyberattack against a government. According to the Financial Times (paywalled), the attack compromised at least 85 accounts and resulted in the theft of more than 2,500 personnel records from Taiwanese systems. Tom's Hardware reports: The campaign reportedly ran for four days at the beginning of July and at times deployed as many as eight autonomous agents in parallel. Dream said the system mapped 21 government systems before compromising user accounts and extracting personnel information. The attackers subsequently expanded their activity to Taiwan's nuclear safety agency, at least seven energy companies, government suppliers, and other government systems.

Dream says it found the evidence inside a 160-megabyte (160MB) online archive that surfaced during its broader tracking of cyberthreat actors. The archive reportedly held 1,395 files showing that the tool was built on two open-source AI agent systems -- Hermes and OpenClaw -- both of which can be downloaded freely and are designed to let large language models carry out multi-step tasks on their own.

Researchers could not determine which underlying model powered the agents, but the data reportedly showed the model's safeguards had been sidestepped by presenting the intrusion as an authorized penetration test rather than a real attack. Of particular concern is that the toolkit for the hack comprised such easily available systems, neither of which was purpose-built for offense. The operators appear to have assembled a capable autonomous tool out of components any developer can pull down and run.

What the researchers describe as the tool's most striking feature was its ability to continuously devise attacks on its own, rather than follow a preprogrammed route. The platform continuously assessed available evidence, ranked possible attack paths, and reprioritized them as circumstances changed. When one technique failed, the tool tasked another agent with searching the internet for information and developing an alternative approach.

Read more of this story at Slashdot.