AI cyberattacks at the gates: viewing the paradigm shift brought by autonomous AI attacks through the Taiwan hacking incident

The security company Dream Security disclosed an autonomous AI attack campaign against Taiwan government agencies, reconstructing from 1,395 work files about 4 days and 12 waves of attacks, with up to 8 AI Agents deployed in a single wave, cracking 85 sets of government personnel account credentials and obtaining at least 2,564 personnel records. The attack chain is divided into 6 stages: target reconnaissance, obtaining SDK, initial intrusion, lateral movement, data theft and supply chain expansion; the system is built on AI Agent architectures such as Hermes and OpenClaw, uses Bayesian posterior probability to assess the success rate of each attack chain, runs at most 5 learning cycles, and recorded 7 misjudgments it overturned on its own. After obtaining account credentials, 84 of the 85 sets could log into another internal system, a success rate of 98.8%, and the scope of the attack extended to government information suppliers, nuclear safety agencies and more than 7 energy companies. The Ministry of Digital Affairs confirmed that in July it detected a hybrid attack by overseas hackers combining AI Agents such as Open Claw, and the Financial Times, citing informed sources, said the victim country was Taiwan.