Taiwan's nuclear safety agency was targeted in what researchers describe as the first 'near-autonomous' AI-enabled cyberattack on a government, according to the Financial Times and The Register. Suspected Chinese operatives built an autonomous hacking tool from open-source AI models — a system that behaved less like an automated script and more like a coordinated cyber team.

The campaign, which Taiwan detected in July, began with intrusions into Taiwanese government systems before expanding to the nuclear regulator. Cybersecurity firm Dream, the Israeli company that uncovered the attack, said the AI agents carried out reconnaissance and break-in attempts simultaneously — coordinating multiple stages of an intrusion in a way that previously would have required a human team directing each step. When blocked, the agents adapted and tried alternative routes without waiting for fresh instructions.

The attack signals a step change in state-backed hacking. Security teams have spent years preparing for attackers who use AI to write phishing emails or analyze stolen data; autonomous agents raise the stakes because they can chain those functions into long, self-directed attack sequences. The choice of target compounds the concern: nuclear agencies sit inside critical infrastructure, where even unsuccessful intrusions can expose sensitive technical details and credentials.

Taiwan's government said the affected agencies successfully fended off the intrusions, and no operational damage has been disclosed. But the incident lands amid growing evidence that frontier AI systems can behave unpredictably during cybersecurity testing, and it sharpens an uncomfortable question for governments and companies alike: how much autonomy should AI be given, when the same capabilities that defend networks can also attack them at machine speed? The Taiwan case suggests that question is no longer theoretical.

Sources