Autonomous Hermes Agent Automates Thai Ministry Breach

Between July 9 and July 13, 2026, threat actors deployed an autonomous AI agent to automate post-exploitation activities during a cyber-espionage operation targeting Thailand’s Ministry of Finance. Discovered by Hunt.io and security researcher Bob Diachenko, the intrusion relied on the open-source Hermes Agent operating entirely without human oversight.

Incident Overview

The operator deployed the Hermes agent on a rented server in Hong Kong to automate the enumeration and traversal of the ministry’s internal network. The attack targeted internal Hadoop infrastructure and mail servers.

Researchers identified active staging infrastructure on an attacker-controlled server (43.246.208[.]207) hosted by TOPIDC. The discovery included three open web directories containing 585 files and approximately 470 MB of attack data. This exposed dataset provided direct access to the automated logs generated by the Hermes agent during the intrusion.

Technical Details

The attack utilized a hybrid approach, combining traditional malware with AI-driven automation. The threat actor leveraged the Hermes AI agent, released in February 2026 by Nous Research, in its documented “YOLO” mode. This configuration disables all human-in-the-loop approval prompts, allowing the agent to execute system commands autonomously.

Once inside the network, the agent ran privilege escalation scripts like LinPEAS, mapped internal hosts, and crawled file systems. It successfully accessed personnel records dating back to 2012. The automated traversal targeted specific vulnerabilities and misconfigurations in Apache Hadoop, Apache Ambari, GlassFish, and HiveServer2, which accepted any password due to a default configuration.

Alongside the AI agent, attackers deployed “Hades,” a previously undocumented, cross-platform Go-based implant. Hades provided persistence, command-and-control (C2), and data exfiltration capabilities on both Windows and Linux systems. The exposed directories also contained suo5 HTTP tunnels, web shells, and scripts with hardcoded stolen credentials.

Attribution and Scope

Hunt.io researchers assess with low-to-moderate confidence that a Sinophone threat actor conducted the operation. This assessment stems from the history of the staging infrastructure, Chinese-language indicators found in the scripts, and the operator’s use of the FOFA search engine for initial reconnaissance.

While the Hermes agent’s logs confirmed deep internal network access and the staging of exfiltration tools, there is currently no evidence of successful large-scale data exfiltration. The Thailand National Cyber Security Agency (NCSA) and ThaiCERT were notified on July 15, 2026, though the Ministry of Finance has not issued a public statement.

If you build defenses against multi-step cyberattacks, this incident demonstrates that internal network monitoring must account for machine-speed traversal. Understanding how AI agents work and applying rate limits or zero-trust validation to internal administrative actions is necessary when attackers can script post-exploitation logic dynamically.