Hermes AI Agent Used in Espionage Attack on Thai Finance Ministry

Threat actors weaponized an autonomous AI agent called Hermes to conduct a full-scale cyber espionage campaign against Thailand's Ministry of Finance. The open-source tool, operated in unrestricted "YOLO mode," allowed the agent to make independent decisions about targeting, data collection, and exfiltration without human oversight, marking a notable escalation in adversarial use of large language model (LLM) frameworks for state-sponsored operations.

According to researchers, the Hermes agent was deployed to autonomously scan internal systems, identify sensitive financial and policy documents, and exfiltrate data through encrypted channels. Running in YOLO mode, the agent skipped typical guardrails and approval gates, executing multi-stage intrusion sequences in a single continuous loop. This approach mirrors tactics attributed to advanced persistent threat (APT) groups, but with the added efficiency and scale that agentic AI provides. Defenders can assess their own external exposure with a port scanner to identify services that an autonomous agent could enumerate and exploit during reconnaissance.

The incident underscores how rapidly the threat landscape is shifting as off-the-shelf AI tools become viable weapons for espionage actors. Unlike traditional malware that follows a predefined playbook, agentic tools like Hermes adapt in real time, rewriting queries, pivoting through networks, and chaining vulnerabilities without operator intervention. Security teams responding to similar threats should prioritize visibility into outbound connections and lateral movement, and a DNS leak test can help surface unintended resolver exposures that might be leveraged for command-and-control traffic. Monitoring for anomalous LLM-driven request patterns is quickly becoming an essential layer of modern detection engineering.

For organizations holding sensitive governmental or financial data, the Hermes campaign is a clear warning that autonomous AI agents are no longer theoretical threats. Defenders should audit their attack surface, enforce strict network segmentation, and adopt zero-trust principles to limit what an unattended agent could reach if it gains a foothold. To stay ahead, security leaders can run a comprehensive privacy checkup and review their organization's digital footprint regularly, ensuring that exposed assets do not become easy entry points for the next AI-driven espionage operation.