Who is suspected of running the Taiwan AI agent attack?

CyberScoop reports that suspected Chinese hackers ran the campaign against Taiwanese government targets using open-source AI models. Dream, the Israeli firm that analysed the intrusion, sets out the operators' tooling and behaviour in detail but frames its report around the multi-agent framework itself rather than naming a state-backed group.

What data was taken in the Taiwan breach?

The Taiwan intrusion produced 2,564 personnel records, 7 SSO client secrets and 6 internal database credentials covering MSSQL, Oracle and Sybase, according to Dream. Investigators also found a 160 MB archive holding 1,395 files. Dream reports that 84 of the 85 cracked credentials authenticated successfully against internal systems.

How autonomous was the attack really?

Dream calls the Taiwan operation near-autonomous rather than fully autonomous. The framework adapted mid-operation without human intervention, but researchers told CyberScoop that building a system that works at this level takes more work than just running a model. People designed the orchestration; the agents did the planning and execution inside it.

Does this mean OpenClaw itself was compromised?

OpenClaw is an open-source agent framework, and Dream's report describes attackers using it as tooling rather than exploiting a flaw in the project. Dream names Hermes and OpenClaw as the frameworks the operators orchestrated, and nothing in the report says either project was breached or backdoored.

What does Dream say defenders should do?

Dream's recommendation is a shift toward AI-native defence: systems that reason probabilistically about which attack path an adversary is most likely to take, rather than waiting to match a known signature. Dream frames this as needed before defensive resources are overwhelmed by attackers who can re-plan between waves.

Dream · 2026-08-12 · major

AI agents breached Taiwan's government — Dream reports 85 accounts cracked

Dream, an Israeli security firm, says attackers used the open-source Hermes and OpenClaw agent frameworks to crack 85 Taiwanese government accounts over four days in July and take 2,564 personnel records, with little human input.

Dream report cover for a multi-agent AI framework used against government entities in Asia

Dream's report describes a four-day agentic intrusion that mapped 21 Taiwanese government systems and reached a nuclear safety agency.

Key specs

Accounts cracked	85
Records taken	2,564

Quick facts

Reported by	Dream (Israeli security firm)
Report date	12 August 2026
Attack window	1–4 July 2026
Frameworks used	Hermes and OpenClaw
Systems mapped	21 connected government systems
Attack waves	12, up to 8 sub-agents each
Other targets	Nuclear safety agency, 7+ energy firms

What is it?

Dream, an Israeli cybersecurity firm, published a report on 12 August 2026 on a multi-agent attack framework run against Taiwanese government targets. The operation took place from 1 to 4 July 2026 and used two open-source agent frameworks, Hermes and OpenClaw. CyberScoop reports that suspected Chinese hackers were behind it.

How does it work?

The framework ran up to eight lettered sub-agents in parallel per wave, and Dream observed Agent A through Agent Q across the campaign. A Bayesian decision engine scored 14 parallel attack chains by posterior probability, so effort went to the highest-value targets first. When a path was blocked, what Dream calls Learning Cycles searched vulnerability databases and GitHub for new techniques, and structured after-action reports carried results from one wave into the next.

Why does it matter?

Near-autonomous intrusion changes the speed defenders face. The framework worked through a dozen waves in four days, re-planning between each one, and spread from government email and IT supply-chain vendors to a nuclear safety agency and energy companies. Dream argues security teams now need an AI-native approach that reasons probabilistically about which attack path is most likely, before defensive resources are overwhelmed.

Who is it for?

security engineers and threat intelligence teams

Frequently asked questions

Who is suspected of running the Taiwan AI agent attack?: CyberScoop reports that suspected Chinese hackers ran the campaign against Taiwanese government targets using open-source AI models. Dream, the Israeli firm that analysed the intrusion, sets out the operators' tooling and behaviour in detail but frames its report around the multi-agent framework itself rather than naming a state-backed group.
What data was taken in the Taiwan breach?: The Taiwan intrusion produced 2,564 personnel records, 7 SSO client secrets and 6 internal database credentials covering MSSQL, Oracle and Sybase, according to Dream. Investigators also found a 160 MB archive holding 1,395 files. Dream reports that 84 of the 85 cracked credentials authenticated successfully against internal systems.
How autonomous was the attack really?: Dream calls the Taiwan operation near-autonomous rather than fully autonomous. The framework adapted mid-operation without human intervention, but researchers told CyberScoop that building a system that works at this level takes more work than just running a model. People designed the orchestration; the agents did the planning and execution inside it.
Does this mean OpenClaw itself was compromised?: OpenClaw is an open-source agent framework, and Dream's report describes attackers using it as tooling rather than exploiting a flaw in the project. Dream names Hermes and OpenClaw as the frameworks the operators orchestrated, and nothing in the report says either project was breached or backdoored.
What does Dream say defenders should do?: Dream's recommendation is a shift toward AI-native defence: systems that reason probabilistically about which attack path an adversary is most likely to take, rather than waiting to match a known signature. Dream frames this as needed before defensive resources are overwhelmed by attackers who can re-plan between waves.