In the wild
Realized harm
Multi-agent framework on Hermes and OpenClaw used in a near-autonomous intrusion into Taiwanese government systems
AIR-2026-0194 AIR record · 2026-07-taiwan-multi-agent-government-intrusion
A real party was affected. This incident caused actual consequences, not a demonstrated capability.
From the supporting source
Suspected Chinese hackers used open-source artificial intelligence models to run a cyberattack against the Taiwanese government in the first publicly known case of an autonomous AI hack hitting a government target, according to research published Wednesday.cyberscoop.com
The record
Disclosed
2026-08-12
Occurred
2026-07-01
Vendor
OpenClaw project; open-weight model providers
Product
Hermes and OpenClaw (abused as attacker infrastructure)
Risk domain
Privacy › Data exfiltration channels
System type
Generative model
Who acted
Agent acted — the agent framework ran the near-autonomous intrusion
Attack surface
Multi-agent system
Autonomy level
Fully autonomous
Initial vector
Misuse of open source agents by threat actor
Impact type
Espionage, Data exfiltration
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI02 – Tool Misuse and Exploitation
- MITRE ATLAS
- AML.T0024 – Exfiltration via AI Inference API
- NIST AI RMF
- MAP 4.2
- EU AI Act
- Article 10
- ISO/IEC
- TR 24028; 23894
- AIUC-1
- A. Data & Privacy
- Scope
- tool misuse
- Mitigation
- Monitor network traffic for anomalous patterns; Restrict egress destinations.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
85 credentials compromised; more than 2,564 personnel records exfiltrated from Taiwanese government systems
Capability abused
Defense & remediation
Recorded defense
government network defences
Guardrail kinds
Unknown
Guardrail outcome
Silent
Taxonomy & references
Curator notes
AI Incident Database incident 1646. Significant because open-weight models and open-source agent frameworks remove the vendor-side kill switch that Anthropic used against GTG-1002. Taiwan confirmed an overseas AI-assisted campaign without attributing it to China.
Supporting sources
Cite as
Enkrypt AI. "Multi-agent framework on Hermes and OpenClaw used in a near-autonomous intrusion into Taiwanese government systems". AIR, AIR-2026-0194. https://agentincident.vercel.app/i/AIR-2026-0194/ (accessed 2026-10-19).