Via livescience.com
A near-autonomous AI system using up to eight parallel sub-agents compromised government infrastructure in Asia, marking the first confirmed cyberattack of its kind against nation-state targets.
A multi-agent AI system just pulled off what cybersecurity researchers have been warning about for years. Over four days in early July 2026, a near-autonomous framework deployed up to eight parallel sub-agents to breach government infrastructure in Asia, extracting more than 2,564 personnel records, 85 cracked credentials, and critical details about internal network architecture.
Dream Research Labs disclosed the operation on August 12, describing it as the first confirmed instance of a multi-agent AI system successfully executing cyber operations against nation-state infrastructure.
How the attack worked #
The operation ran from July 1 to July 4, 2026, targeting government systems with Taiwan as the likely focal point. The framework was built on components from the Hermes and OpenClaw agent platforms, two tools that, when combined, gave the system a disturbingly wide range of autonomous capabilities.
Up to eight sub-agents, designated A through Q, worked in parallel to handle different phases of the intrusion. Some performed reconnaissance. Others focused on credential cracking. Others handled vulnerability exploitation and data exfiltration. The whole thing operated with minimal human oversight.
The framework automated the decision-making process, using autonomous learning cycles and Bayesian probabilistic scoring to prioritize which vulnerabilities to exploit and which targets to pursue next.
The operation produced 1,395 files and an operational archive exceeding 160 MB. Beyond personnel records and credentials, the attackers gained access to interconnected systems including those belonging to supply-chain vendors and energy-sector entities, with nuclear safety systems among the compromised infrastructure.
Attribution and the language trail #
Code-switching between Simplified and Traditional Chinese characters appeared throughout the system’s operational logs. That pattern, combined with other forensic indicators, pointed toward a Chinese-language operator.
Dream Research Labs noted that affected organizations received notifications before the public disclosure, following responsible disclosure protocols.
Why this changes the threat landscape #
Previously, a four-day operation of this complexity would have required a team of skilled hackers coordinating across multiple specialties: network reconnaissance, credential exploitation, lateral movement, data exfiltration. This framework compressed all of that into an autonomous pipeline. The sub-agents handled specialization. The Bayesian scoring system handled prioritization. The learning cycles handled adaptation when initial approaches failed.
The involvement of energy-sector entities and nuclear safety systems adds another layer of concern, as unauthorized access to operational technology networks in the energy sector has been a top-tier worry for national security officials.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our