Researchers say hackers with suspected links to China carried out what is described as the first observed end-to-end autonomous cyberattack against a government target. The campaign targeted government systems in Taiwan and lasted approximately four days at the beginning of July.
According to researchers at the Israeli cybersecurity company Dream, the attackers used an autonomous hacking platform built on open-source AI-agent frameworks, identifying the system as using Hermes and OpenClaw. During the attack, up to eight autonomous AI agents were deployed simultaneously, mapping 21 government systems and compromising at least 85 user accounts.
Government systems mapped during the four-day campaign.
The target list expanded to include Taiwan's nuclear safety agency and at least seven energy companies. Researchers identified a 160-megabyte archive containing 1,395 files used to uncover the system, and more than 2,500 personnel records were reportedly stolen from the systems.
The underlying large language model powering the agents could not be definitively determined. The attackers bypassed the AI model's safeguards by presenting the intrusion as an authorized penetration test, and the system's ability to continuously devise attacks on its own—rather than following a preprogrammed route—was described by researchers at Dream as its most striking feature.
Researchers noted that the operators' internal communications used Simplified Chinese, suggesting a high probability of a connection to China. In contrast, the data extracted from the target systems was written in Traditional Chinese.
The Financial Times reported on August 12 that hackers with suspected links to China used AI to run the first-ever end-to-end autonomous cyberattack on Taiwan's government, according to researchers at Dream. According to Taiwan's security agencies, the daily average of cyberattacks suspected to be from China reached approximately 2.6 million in 2025.
