AI Agent Runs Espionage Attack on Thailand Finance Ministry

Threat actors breached Thailand's Ministry of Finance using an open-source AI agent running in fully autonomous mode—one of the first documented cases of AI-driven espionage against a government financial institution.

Hunt.io researchers uncovered the operation after discovering three exposed directories on a Hong Kong-based server containing 585 files totaling approximately 470 MB of exploit code, web shells, and AI operation logs. The attackers had deployed the Hermes AI agent in "YOLO mode," which permits execution of dangerous commands without human approval.

Autonomous Post-Exploitation

Once inside the network, the Hermes agent handled system enumeration, privilege escalation, and network reconnaissance without human supervision. The agent cataloged PDF, DOC, and XLS files from the Office of Permanent Secretary for Finance; investigators found no evidence that documents were exfiltrated.

Alongside the AI agent, attackers deployed a custom Go-based implant called Hades, which provided interactive remote shell access, persistence mechanisms, and SOCKS proxying capabilities for Windows and Linux systems.

Attribution Remains Uncertain

Hunt.io assessed with low-to-medium confidence that the threat actor is Chinese-speaking, citing the Hong Kong server infrastructure and previous links to ShadowPad controllers. The agency noted Chinese-language indicators in the tooling but stopped short of definitive attribution.

Growing Trend of AI-Assisted Attacks

The Thailand incident fits a broader pattern of threat actors offloading routine offensive operations to LLM-powered agents. Recent months have seen the JadePuffer ransomware operation employ similar automation, and Hunt.io has tracked suspected Chinese operators using Claude Code and DeepSeek to target financial systems across four countries.

Autonomous agents let attackers scale post-exploitation activities, maintaining high-tempo reconnaissance and lateral movement without constant human oversight. But the volume of logs these agents generate creates a larger forensic footprint.

"The same tooling that speeds these operations up also produces a paper trail, and if it's left exposed it becomes a direct window into how the intrusion ran," said Esteban Borges, head of research at Hunt.io.

In this case, poor operational security—leaving directories publicly accessible—enabled discovery of the entire operation. The exposed logs give researchers rare visibility into how AI agents function during active intrusions.