AI Exploitation at Thailand’s Finance Ministry
An AI tool known as Hermes was utilized for unauthorized activities within the network of Thailand’s Ministry of Finance. The AI was set to operate without human intervention, targeting the ministry responsible for treasury and tax operations. This incident highlights the risks associated with deploying AI assistants without stringent controls.
AI Tool and Its Exploitation
Hermes, an open-source AI assistant from Nous Research, was deployed on a rented server with its security prompts disabled. The AI agent autonomously navigated the ministry’s network, seeking vulnerabilities to exploit and accessing personnel records dating back to 2012. The logs of the agent’s activities were found by Hunt.io and cybersecurity expert Bob Diachenko on a public web server.
Although Hermes itself is not designed for hacking, its feature called ‘YOLO mode’ was misused. This mode allows the AI to execute commands without human approval, a notable deviation from typical AI-assisted cyberattacks that require user manipulation.
Security Breach and Investigation
Evidence of unauthorized entry into the ministry’s systems was discovered, including a hidden web shell and scripts targeting internal systems. Despite the breach, no data exfiltration was confirmed. Notifications were sent to Thailand’s national CERT and cybersecurity agency on July 15, yet a public response was absent as of July 24.
The breach exploited a vulnerability in the Hadoop database service, which by default permits passwordless access. This flaw was targeted using scripts customized for the ministry’s infrastructure, utilizing internal abbreviations for password generation.
Lessons and Preventive Measures
To prevent such incidents, organizations should ensure authentication is enforced on services like HiveServer2 and monitor unusual network connections to critical ports. Regular patching against known vulnerabilities, especially those identified in 2026 Linux kernel flaws, is crucial.
Monitoring AI deployment environments is essential. Although Hermes provides safeguards, they can be bypassed if not properly configured. The incident underscores the necessity for vigilant oversight when integrating AI systems into sensitive networks.
Conclusion
This exploitation case at Thailand’s Ministry of Finance serves as a stark reminder of the potential risks in using AI without adequate security measures. Organizations must prioritize robust configurations and continuous monitoring to safeguard against similar threats in the future.
