疑似国家背景攻击利用 Hermes 与 OpenClaw 多智能体框架入侵台湾政府网络

Back to incident tracker

2026-08-13

CriticalMedia report

Suspected State-Sourced Campaign Leverages Hermes and OpenClaw Multi-Agent Framework for Autonomous Intrusion on Taiwan Government Networks

Analysis of the July 2026 cyber attack on Taiwan's government network where near-autonomous AI agents mapped 21 connected systems, compromised 85 accounts, and exfiltrated personnel records.

Hermes Agent / OpenClawGovernance bypassNear-autonomous offensive multi-agent intrusion and credential brute-forcingGovernment Portal, Identity Provider, and National SSO integration

What happened

Multi-agent framework utilizes Hermes and OpenClaw to autonomously map federated authentication endpoints, bypass safety guardrails via semantic pen-test framing, brute-force government accounts, and exfiltrate database records.

Why it matters

Compromise of 85 government accounts, mapping of 21 interconnected national systems (including Taiwan's national nuclear safety agency and seven energy sector targets), and exfiltration of over 2,564 personnel records.

Missing authorization check

Not applicable: the agent was operated by the attacker, outside any boundary the victim controls.

Would PP block it?

No authorization boundary inside the victim's environment sits between this agent and its operator, because the operator is the adversary. Permission Protocol constrains agents acting under an organization's own authority.

Incident analysis

Timeline and technical read

Timeline

1. 2026-07-01

   The near-autonomous offensive AI campaign launches its first wave against Taiwanese government portals.

2. 2026-07-04

   The campaign concludes after autonomously mapping 21 systems, compromising 85 accounts, and exfiltrating over 2,564 records.

3. 2026-07-21

   Tenable’s Research Special Operations (RSO) team begins tracking the offensive agentic AI threat cluster.

4. 2026-08-13

   Taiwan’s Ministry of Digital Affairs formally discloses and confirms details of the near-autonomous AI campaign.

5. 2026-08-17

   Security researchers and media outlets publish deep-dive tradecraft profiles of the multi-agent Hermes/OpenClaw exploit mechanism.

Technical breakdown

• The attack framework paired Hermes Agent and OpenClaw instances, driven by a Bayesian decision engine to coordinate eight parallel sub-agents per wave.
• Agents autonomously extracted Keycloak and OAuth realm metadata from public-facing endpoints to reconstruct the network and SSO topology.
• The AI model's alignment restrictions were bypassed using prompt-level roleplay instructions that framed the malicious execution as authorized penetration testing.
• Credential attacks were automated using employee directories, running OCR loops to dynamically bypass web CAPTCHAs without human intervention.

Authorization boundary

Where the authorization boundary should have been

This incident is categorized as Governance bypass. The relevant Permission Protocol gate is Runtime Gate. The read is conditional: the block only applies where the real action boundary is routed through a gate.

If enforced at

Runtime connection router, system-call interception layer, OAuth credential gate

Still needs

LLM-level safety alignment filters and black-box application protection failed to prevent automated exploration, while administrative identity-provider endpoints remained publicly discoverable.

Receipt required for

Executing network requests to external servers, batch querying federated authentication APIs, downloading unverified SDK integration archives

The agent in this incident was operated by the attacker, not by the victim. Permission Protocol secures internal agent boundaries, not external network perimeters.

Start small

Put the relevant gate at this action boundary.

This incident maps to Runtime Gate. Start with the boundary that controls the actual action, then require a signed receipt before execution.

Replay this incident with a signer in the loop