Taiwan detected an AI-assisted cyberattack against government agencies in July, with the Ministry of Digital Affairs reporting that affected bodies completed incident handling. The Register and other outlets report that a near-autonomous system using open-source agents reached government, nuclear-safety, and energy-sector targets. Taiwan did not publicly attribute the incident to a specific actor.
Taiwan detected an AI-assisted cyberattack against government agencies in July, and the Ministry of Digital Affairs said the affected bodies had completed their handling of the incident. Reuters reported that the ministry described the activity as originating overseas and combining manual operations with AI agent-assisted attacks, including OpenClaw.
The case has drawn attention because reporting by The Register, The Guardian, and TechRadar describes a more extensive intrusion involving open-source AI agents and targets in Taiwan's government and critical-infrastructure ecosystem. Taiwan's public statement did not identify an attacker or confirm the broader technical claims attributed in those reports to Dream, an Israeli AI and cyberdefense company.
What the reporting describes
According to The Register, suspected Chinese operators conducted the activity during the first four days of July through 12 attack waves. The outlet reported that the system, built on the Hermes and OpenClaw agents, was "near-autonomous" and used as many as eight sub-agents assigned to separate targets and techniques.
The Register reported that the intrusion reached a Taiwanese government website and then a government email system, the nuclear safety agency, IT supply-chain vendors, and at least seven energy companies. TechRadar, citing reporting based on Dream's research, reported that the attackers compromised 85 government accounts and extracted more than 2,500 personnel records before expanding to the nuclear-safety agency and energy companies.
The Guardian similarly reported that Dream characterized the tool as an autonomous hacking system behaving like a coordinated cyber team. It reported that Dream did not assign the activity to a named group, though the company cited Simplified Chinese used in internal communications as a reason it assessed a high probability of a China connection. Reuters noted that Taiwan's ministry did not mention China in its statement.
That distinction matters. The confirmed public account from Taiwan describes AI-assisted activity and successful handling, while several reports characterize the campaign as near-autonomous or fully autonomous based on Dream's assessment. These labels should not be treated as interchangeable without technical evidence on operator approval points, tool permissions, and the degree to which agents independently selected and executed actions.
Critical-infrastructure risk
At Black Hat, FBI Cyber Division Assistant Director Brett Leatherman told The Register that threats to critical infrastructure were the bureau's central concern. "That is where cyber becomes kinetic," he said, naming water and wastewater systems, the electric grid, and high-frequency-trading and financial networks as systems whose compromised integrity could affect communities and national security.
Tom Kellermann, TrendAI vice president of AI security and threat research, told The Register that "there is a clear and present danger" from autonomous AI-enabled destructive attacks. His comments are an assessment of future risk, not evidence that the Taiwan incident caused physical damage. None of the retrieved reports states that the July intrusion produced a kinetic effect or disrupted energy operations.
Taiwan's Ministry of Digital Affairs said it issued warning alerts through its National Institute of Cyber Security beginning July 20 and strengthened system monitoring and protective guidance across agencies, Reuters reported.
Implications for defenders
Anthropic's June research offers a broader data point on AI-enabled abuse. The company reported analyzing 832 Claude accounts associated with malicious cyber activity from March 2025 through March 2026, all of which it had banned for policy violations. That research does not document the Taiwan campaign, but it illustrates that vendors are observing AI use across real-world cyber operations.
For security and ML practitioners, the Taiwan reporting reinforces an industry-wide pattern: agentic systems can lower the operational cost of recon, vulnerability chaining, adaptation, and parallel task execution when connected to tools and credentials. Defensive evaluations therefore need to test not only model outputs, but also the permissions, network reach, logging, approval gates, and recovery controls surrounding autonomous or semi-autonomous workflows.
Key Points #
- 1Taiwan confirmed an AI-assisted July intrusion, while external reporting describes broader critical-infrastructure targeting by a near-autonomous agent system.
- 2Reports distinguish official attribution from Dream's China-link assessment, leaving the responsible actor unconfirmed by Taiwan's government.
- 3Comparable agentic threats increase the importance of testing tool permissions, approval gates, telemetry, and containment rather than model behavior alone.
Scoring Rationale #
The incident is a significant reported example of AI agents being used against government and critical-infrastructure targets. It is especially relevant to security and ML teams deploying tool-using agents, although key technical details and attribution remain based on external reporting rather than Taiwan's public statement.
Sources #
Primary source and supporting public references used for this report.
View 5 more sources #
Taiwan says it was targeted last month in AI-driven hacking campaignreuters.comTaiwan says it was hit by ‘abnormal’ AI-assisted cyber-attacktheguardian.comChina-linked hackers hit Taiwan in unprecedented ‘autonomous’ AI cyber attackft.comAutonomous AI attacks pose 'clear and present danger' to critical infrastructuretheregister.comWorld-first autonomous ‘end-to-end’ AI attack against Taiwan tied to Chinese hackers — and the scariest part is that it was fully open sourcetechradar.com
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.