AI Agents Breach Taiwan Government Systems in Four Days, Leaving 12 Attack Waves

One missing piece of first-hand evidence determines how autonomous this incident really was. Dream has not released the complete operational workspace it recovered, and Taiwan has not published packet-level logs. If a later audit shows several attack paths running at the same time and reprioritizing after blocks without intervening operator commands, the “near-autonomous” description gains support; if command timing follows a human work rhythm, this should be classified as a human-led attack accelerated by AI.

Dream published its investigation on August 12, 2026. The company said its researchers recovered more than 160 MB from a multi-agent attack workspace aimed at government entities in Asia in early July 2026. The archive covered 12 attack waves from July 1 through July 4 and contained 1,395 files. Built around the Hermes and OpenClaw agent frameworks, the system dispatched as many as eight sub-agents in parallel during a wave, assigned them separate reconnaissance and exploitation tasks, and fed their results into the next planning cycle.

One identity chain connected 21 systems

According to Dream, the agents began with JavaScript from a government portal and extracted API paths, an OAuth client ID, and Keycloak configuration. They then mapped 21 connected systems and six single-sign-on sub-realms. Unauthenticated APIs exposed account data that supported password spraying, which produced 85 cracked credentials. Of those accounts, 84 authenticated to an internal system through an SSO bridge, a reported success rate of 98.8 percent. The workspace also recorded an API accepting unsigned JWTs, production debug endpoints that returned authenticated sessions, and a web-shell upload whose execution was stopped by a second authentication layer.

Dream reported that the operation obtained 2,564+ personnel records, seven rotated SSO client secrets, and six database credentials, while also scanning a nuclear-safety agency, a government email system, and energy companies. These figures come from Dream alone. When Reuters asked for the underlying material, the company declined to share it or identify the affected government. Dream linked the operator to a Chinese-language environment because internal reports used Simplified Chinese while target-facing analysis used Traditional Chinese, but language is not sufficient by itself for attribution to a state.

Taiwan confirms an overseas attack, not every reported figure

Taiwan’s Ministry of Digital Affairs said on August 13 that monitoring teams had detected an abnormal attack against government agencies in July. The National Institute of Cyber Security began issuing alerts on July 20, and the ministry described the activity as an overseas operation mixing manual work with AI agents such as Open Claw. It said the affected bodies had completed their handling of the incident. The statement did not name China, identify the agencies, or confirm the figures for credentials and personnel records, so it corroborates an AI-assisted attack on Taiwan but not every item reconstructed from Dream’s archive.

Reuters also noted that a person still had to choose the target, define the objective, and direct the system. Dream’s evidence nevertheless describes parallel execution, reprioritization after failure, and a verification process that removed mistakes. In one example, the framework first treated a 21-second response delay as evidence of SQL injection, retested it, and concluded that an email-server timeout caused the delay; its final record listed seven discarded false positives. Over the next three to six months, the most useful measures will be time from first reconnaissance to credential use, the number of concurrent attack paths, gaps between operator commands, and the minutes needed to move to a new endpoint after one route is blocked.

Sources:

• Dream: Inside a Multi-Agent AI Framework Used to Compromise Government Entities in Asia
• Reuters: Taiwan says it was targeted last month in AI-driven hacking campaign
• Central News Agency: Overseas hackers used AI agents to attack government websites