For years, the idea of AI carrying out a cyberattack sounded like it was made up. A model could find vulnerabilities or write malicious code, but there was still a person deciding what happened next.
But this became much harder in the advent of AI, and especially in July, when AI agents were used in a four-day operation against Taiwanese government systems and, according to researchers, kept the attack moving by finding new routes when earlier ones failed.
Dream, the Israeli cybersecurity company that uncovered the operation, says the agents mapped 21 government systems, compromised at least 85 accounts, and extracted more than 2,500 personnel records. Taiwan has since confirmed that its government agencies were targeted in an overseas attack that combined conventional hacking with AI-agent assistance.
This attack is significant largely because the AI system reportedly handled much of the operation itself after it was set in motion, with Dream describing it as the first end-to-end autonomous cyberattack against a government target that it had observed. However, Taiwan’s Ministry of Digital Affairs said the wider campaign used a hybrid approach involving both human-controlled operations and AI agents such as OpenClaw.
The Attack Ran for Four Days
According to Dream’s findings, the operation ran from July 1 to July 4 and used up to eight AI agents at the same time. The agents mapped 21 government systems, searched for weaknesses, and adjusted their approach when an attack path failed.
Researchers discovered evidence of the operation in a 160MB archive containing 1,395 files, and the material showed that the attackers built their framework using two publicly available AI agent systems called Hermes and OpenClaw. However, Dream could not determine which underlying AI model powered the agents.
Still, the agents were able to identify exposed systems and authentication weaknesses before moving deeper into the targeted network. Of the 85 accounts the attackers compromised, 84 could then be used to access internal systems through the organisation’s SSO infrastructure. The compromised accounts gave access to internal dashboards, equipment management interfaces, and personnel information.
The operation also extracted more than 2,500 personnel records, seven SSO client secrets, six internal database credentials, and other internal network information. The attackers later expanded their activity to Taiwan’s nuclear safety agency, government IT suppliers, a government email system, and at least seven energy companies.
The AI Changed Tactics on Its Own
One of the most important findings was how the agents responded when their first attempts failed.
Dream said the system continuously ranked possible attack paths based on the information it collected. When one route stopped working, another agent searched online sources for information about vulnerabilities and potential techniques before developing another approach.
The framework also used what researchers described as “learning cycles,” allowing the agents to search vulnerability databases, GitHub repositories, and security research for information relevant to the systems they were targeting.
The underlying model’s safety controls were reportedly bypassed by presenting the activity as an authorised penetration test.
Taiwan Has Confirmed the Attack
Taiwan’s Ministry of Digital Affairs said its cybersecurity monitoring units detected the abnormal attacks in July. From July 20, the National Institute of Cyber Security issued a series of warnings while authorities investigated the source, methods and impact.
The ministry said the attacks showed clear signs of an overseas source and involved a combination of conventional hacking and AI-assisted techniques. It also said affected agencies had completed their response and that the government has introduced new protective guidelines and strengthened monitoring.
Dream did not identify a specific hacking group or officially attribute the operation to China. Researchers did, however, find Simplified Chinese in communications linked to the operation, which they said suggested a likely connection to China.
The incident gives cybersecurity teams something more concrete to respond to than theories. AI agents have now been documented performing reconnaissance, finding weaknesses, adapting their methods, and accessing real government systems over several days. The technology used to assemble the operation was also publicly available.
This combination is what makes the Taiwan attack a major warning for governments and organisations already relying on increasingly autonomous AI systems.
