Cybersecurity researchers have identified what they describe as the first "near-autonomous" AI-powered cyberattack targeting a government entity, specifically the Taiwanese government. The attack, which utilized open-source AI models, resulted in the exfiltration of over 2,500 personnel records and other data.
The Israeli cybersecurity firm Dream, which uncovered the operation, detailed that the attack framework was designed to adapt and expand mid-operation without direct human intervention. This included "Learning Cycles" where the AI system autonomously searched vulnerability databases, GitHub repositories, and security research publications for techniques relevant to the target's infrastructure.
The attackers did not limit their scope to primary targets. The operation expanded to include government IT supply chain vendors, a nuclear safety agency, a government email system, and more than seven energy sector companies. The AI system scanned these entities in parallel for misconfigurations, exposed administrative interfaces, and exploitable vulnerabilities, learning from its mistakes as the campaign progressed.
Dream discovered the operation through an online archive containing 160 megabytes and nearly 1,400 files, which provided evidence of a multi-agent AI system achieving confirmed compromises against state infrastructure. The attackers reportedly used two popular open-source AI frameworks, Hermes and OpenClaw, for the Taiwan operation. They allegedly bypassed safety guardrails by presenting the activity as authorized penetration testing.
While described as "near-autonomous," the researchers noted that such sophisticated AI-powered offensive operations still require significant human effort in their setup and refinement. This includes careful adjustment to specific tasks, optimization of agent coordination, and fine-tuning of decision logic, exemplified by the framework's Bayesian prioritization, self-correction loops, and adaptive research cycles.
This incident follows previous reports of autonomous AI-powered cyber campaigns. Last fall, another cybersecurity firm, Anthropic, reported stopping what it identified as the first autonomous cyber espionage campaign, though that operation also required substantial human involvement. The current attack against Taiwan marks the first publicly known instance of an autonomous AI hack targeting a government entity.
