Taiwan Confirms AI Agents Are Now Part Of Real Cyberattacks

Taiwan has now officially confirmed what cybersecurity researchers had warned about this week: AI agents are no longer only a lab concern or a dramatic conference demo. They are being used in real cyberattacks against government systems.

In an August 13 statement, Taiwan’s Administration for Cyber Security said its monitoring units detected an abnormal attack against government agencies in July. The agency said warning alerts began on July 20 through the National Institute of Cyber Security, and that the relevant sources, methods and scope of impact had been investigated, with affected units completing response work.

The key detail is the attack method. Taiwan said the incident showed clear signs of an overseas source and involved a hybrid mode that combined hacker operations with AI-agent-assisted attacks, including tools such as Open Claw. The agency said AI agents can quickly combine multiple attack methods and use backup or test systems as stepping stones, making attacks faster, cheaper and easier to scale.

That official language matters because it moves the story from private-sector warning to government-confirmed reality. The earlier reporting from FT and security firm Dream described a coordinated AI-agent attack against Taiwanese government systems. Taiwan’s statement does not publicly blame China, but it confirms the broader pattern: overseas hackers used AI-agent support in an abnormal campaign against state systems.

The government also said it has established protective guidelines for AI-derived cybersecurity threats, strengthened monitoring across agencies and improved cross-agency intelligence sharing. In other words, Taiwan is not treating this as a one-off technical incident. It is treating AI-assisted cyber operations as a new category of national defence problem.

This is exactly where the cyber debate has been heading. AI agents can chain tasks, gather information, test routes, retry failures and coordinate actions faster than a human operator doing everything manually. They do not remove the human from the operation entirely. A human still chooses the target and objective. But agents can lower the cost and increase the tempo of attacks.

We wrote earlier that AI agents are entering real cyberwarfare. Taiwan’s confirmation strengthens that point. It also links to the wider frontier-model security concern, including research showing that frontier AI models can leak hidden reasoning traces and OpenAI’s decision to restrict stronger cyber capabilities to vetted defenders through Daybreak.

For governments, the lesson is blunt. Defensive teams now have to assume that attackers will use agents for reconnaissance, vulnerability testing, credential abuse, lateral movement and data collection. The old assumption that automation only helps with simple scanning is no longer enough.

For companies, the Taiwan case should be read as an early warning. If AI-agent attacks are being tested against government systems today, they will eventually be used against banks, telecoms, energy firms, cloud providers, hospitals and ordinary businesses. Attackers will not wait for AI policy to mature before using whatever works.

The AI cyber race is therefore becoming a speed race. Defenders need agentic tools too, but they also need better identity controls, segmentation, logging, API security and rapid response. AI may help defenders move faster, but only if their systems are ready for it.

Taiwan’s statement is careful and sober, but the implication is not small. The internet is entering a period where AI agents will operate on both sides of the security line. The countries and companies that understand that early will have a better chance of absorbing the next wave of attacks before it becomes normal.