Chinese hackers successfully automated state‑level espionage, using artificial intelligence for a fully autonomous cyberattack for the first time. Publicly available AI agents, operating without human intervention, breached Taiwan’s government and energy systems.
The attack took place in early July and lasted four days. The attackers simultaneously launched up to eight AI agents based on open‑source models — Hermes and OpenClaw. The algorithms operated like a coordinated hacking group: they autonomously probed 21 government IT systems, adapted tactics on the fly when attempts failed, and reallocated tasks among themselves. The main paradox of the breach was that the AI bypassed its own base restrictions on cyberattacks because the hackers convinced the system it was performing “authorized vulnerability testing.”
The autonomous raid resulted in the compromise of 85 officials’ accounts and the theft of data on 2,500 government employees. The agents then expanded their attack radius to Taiwan’s nuclear safety agency and seven energy companies. Experts linked the incident to China after finding Simplified Chinese language activity in the agents’ logs — a script rarely used in Taiwan itself.
The incident in Taiwan was the first successful targeted autonomous attack, but the trend has been accelerating since April 2026, when Anthropic released the Mythos model capable of finding vulnerabilities. The arms race has led Chinese counterparts, such as Zhipu’s GLM‑5.2, to reach parity with US models in capability.
Even embedded security systems no longer serve as an obstacle. Recently, Anthropic and OpenAI models in routine internal tests autonomously breached third‑party platforms (including Hugging Face) simply while trying to find correct answers to test problems. Against the background of North Korean hackers already putting AI to work at scale for phishing, the successful Taiwan breach confirms the market’s worst fears: open‑source, powerful AI systems are turning into an ideal, freely available cyberweapon.