Open-Source Hermes AI Agent Used to Automate Cyberattack on Thai Finance Ministry
Key Takeaways
- ▸Open-source Hermes AI agent operated in unattended 'YOLO' mode to automate cyberattack tasks without human approval
- ▸Attack infrastructure exposed 585 files (470 MB) including exploit code, web shells, and Hermes logs targeting Thai Finance Ministry systems
- ▸AI agent was used to enumerate systems, scan for privilege escalation opportunities, and analyze kernel vulnerabilities
Summary
Threat actors leveraged the open-source Hermes AI agent to automate post-exploitation activities during an alleged breach of Thailand's Ministry of Finance, according to threat intelligence firm Hunt.io and security researcher Bob Diachenko. The attackers operated Hermes in unattended 'YOLO' mode—a feature that removes safety prompts normally requiring human approval before executing dangerous commands. This automation allowed the threat actors to conduct reconnaissance, privilege escalation attempts, and system enumeration without manual intervention at each step, significantly accelerating the attack timeline.
The attack was uncovered when researchers discovered exposed directories on a server hosted in Hong Kong containing 585 files totaling approximately 470 MB. The recovered artifacts included exploit code, web shells, stolen credentials, command logs, and detailed records of Hermes operations targeting Ministry of Finance infrastructure—including references to Hadoop systems, Apache Ambari management platforms, GlassFish consoles, and mail servers. Hermes logs revealed the agent was deployed to find privilege escalation vectors, scan for kernel vulnerabilities, enumerate services, and traverse file systems.
This incident exposes a significant security risk: open-source AI agents designed for legitimate automation can become powerful tools for cybercriminals when their safety guardrails are disabled. While the Thai Ministry of Finance has not confirmed a successful breach, the recovered evidence suggests attackers gained meaningful access to internal systems, raising questions about the safety defaults of widely-distributed AI agent tools.
- Incident highlights critical security risks of deploying AI agents with disabled safety guardrails in adversarial environments
Editorial Opinion
This incident reveals a troubling design flaw in how we build automation tools: the 'YOLO' mode that prioritizes unattended execution over safety becomes a dangerous liability when tools spread widely and interact directly with system infrastructure. As AI agents grow more capable and prevalent in enterprise environments, the security community must establish mandatory safety defaults rather than treating human oversight as an optional toggle. Open-source projects providing powerful automation capabilities bear a responsibility to make responsible deployment the default, not an afterthought.
