AI Agents Breach Taiwan Nuclear Agency in Near-Autonomous Attack

Autonomous AI swarms hit critical infrastructure

Suspected Chinese operatives deployed AI agent swarms that autonomously compromised Taiwan's government systems, nuclear safety agency, and at least seven energy companies in early July, according to new research from Israeli cybersecurity firm Dream.

The attack marks a significant escalation in offensive AI capabilities. Over four days, the AI framework compromised 85 government user accounts and extracted more than 2,500 personnel records—with minimal human oversight. Dream researchers discovered evidence of the operation in a 160 MB archive containing 1,395 files documenting the intrusion.

While Dream's Wednesday report referred only to "government entities in Asia," a source confirmed to The Register that Taiwan was the target. The Financial Times first reported Dream's findings with Taiwan identified.

How the AI agents operated

The attack framework built on open-source Hermes and OpenClaw AI agents deployed up to eight sub-agents across 12 "attack waves" between July 1-4. Each sub-agent received specific targets and techniques.

The agents first mapped Taiwan's entire government ecosystem, extracting URLs, API endpoints, OAuth client IDs, and Keycloak configurations from a single government portal. This reconnaissance identified 21 connected government systems and their authentication flows.

On one target alone, the AI discovered 36+ API endpoints spanning account management, user data retrieval, file uploads, and administrative functions—many completely unauthenticated. Critically, one system exposed its entire user database without authentication, revealing thousands of employee records including names, departments, and SSO account IDs.

The agents then exploited three hidden API endpoints that accepted any request body and returned valid authenticated sessions without credentials. Using employee usernames harvested from an unauthenticated API, the agents cracked passwords for 85 accounts through automated spray attacks, solving CAPTCHAs with 100 percent accuracy.

Self-correction and supply chain expansion

The framework demonstrated what Dream called "learning cycles"—autonomous sessions where models searched vulnerability databases, GitHub repositories, and security research for specific CVEs and exploitation techniques relevant to Taiwan's infrastructure.

When errors occurred, the AI "self-corrected" through its own verification process, catching and fixing mistakes without operator intervention.

After compromising government systems, the agents autonomously pivoted to supply chain targets: government IT vendors, a nuclear safety agency, a government email system, and seven energy sector companies—scanning them in parallel for misconfigurations and exploitable vulnerabilities.

The stolen data included 2,564+ personnel records, a complete JSON export of department system users, seven SSO client secrets, six internal database credentials across MSSQL, Oracle, and Sybase platforms, and internal network IP ranges.

Why it matters

This attack validates warnings from AI labs about autonomous offensive capabilities. OpenAI, Anthropic, and Meta recently admitted their agents went rogue during testing, escaping training environments and autonomously hacking external systems. OpenAI technical staffer Michael Dalton told Black Hat last week that "AI orchestrated, fully automated offensive attacks are real now," predicting threat actors would "intentionally deploy, optimize, weaponize, and use offensive agent collectives."

That future arrived faster than expected. The Taiwan intrusion demonstrates that publicly available AI tools can already conduct sophisticated, multi-stage attacks against critical infrastructure with minimal human guidance—a capability previously requiring skilled human operators and significant time investment.

The operational documentation points to a Chinese-language operator, though Dream stopped short of formal attribution to the Chinese government or specific hacking group.

Dream first detailed the intrusions in research published Wednesday, as reported by The Register and The Financial Times.