Hackers Use Open-Source AI Agents to Breach Taiwan Government Systems

Open-source AI agents are designed to plan and execute complex, multi-step tasks with limited human supervision, but the same capabilities can be repurposed for cyberattacks. Israeli cybersecurity company Dream said tools including Hermes and OpenClaw were used to build an autonomous intrusion system, highlighting how agentic AI can accelerate reconnaissance, exploitation and data theft while challenging conventional safety-alignment and defensive controls.

Dream said the attackers penetrated 21 Taiwanese government systems over four days and stole more than 2,500 personnel records. The campaign later widened to include the Nuclear Safety Commission, government IT supply-chain targets and energy companies. The breadth and coordination of the operation showed that AI-enabled attack systems can move between targets and sustain intrusion activity with a high degree of autonomy, security experts warned.