Chinese hackers use AI to launch autonomous attack on Taiwan
What's the story
A group of Chinese hackers has reportedly used open-source artificial intelligence (AI) agents to launch a "near-autonomous" cyberattack on the Taiwanese government. The operation was so effective that it cracked 85 government accounts within an hour through password guessing. It also stole thousands of records and exposed vulnerabilities in a government web application, according to Israeli cybersecurity provider Dream.
Attack method
Hackers developed software framework using open-source AI agents
The cyberattack was carried out using a hacker-developed software framework that leverages free and open-source AI agents, Hermes and OpenClaw.
These agents can work autonomously on a computer.
The framework employed at least eight "sub-agents," each assigned specific tasks such as vulnerability testing, password guessing, and supply-chain reconnaissance.
Security breach
AI agents bypassed safety guardrails by framing instructions
The AI agents are said to have bypassed their own safety guardrails by framing the instructions as "authorized penetration testing." This is a form of security testing usually conducted by cybersecurity professionals.
Researchers at Dream discovered this threat last month when they found an autonomous AI attack framework actively conducting intrusion campaigns against government entities in Asia, specifically Taiwan.
Target identification
Simplified Chinese was used in internal status reports
The computer code used to breach the systems revealed Simplified Chinese was used in internal status reports while Traditional Chinese appeared in target-facing analysis.
This strongly indicates that Taiwan was the intended target, where Traditional Chinese is predominantly used.
The investigation found that the attack had stolen over 2,564 personnel records along with a complete user database.
Target expansion
Hackers probed government IT supply-chain vendors, nuclear safety agency
The hackers also expanded their reach beyond the initial targets, probing government IT supply-chain vendors, a nuclear safety agency, and a government email system. They even targeted over seven energy-sector companies.
These systems were scanned simultaneously for misconfigurations, exposed administrative interfaces, and exploitable vulnerabilities.
The incident highlights the growing risks posed by autonomous AI systems being used for malicious cyber operations.