AI Agent Used in Cyber Espionage Against Thailand's Ministry of Finance

A recent cyber attack on Thailand's Ministry of Finance involved the use of the open-source Hermes AI agent, which was configured to run in an unattended mode known as "YOLO." By disabling safety prompts, the attacker enabled the agent to autonomously carry out reconnaissance, enumerate internal networks, and search for privilege escalation paths using tools such as LinPEAS. This operation came to light when researchers discovered exposed staging servers that contained logs from the agent, custom scripts, and a previously undocumented implant referred to as "Hades."

The initial access point for the attacker remains unknown, but they subsequently deployed a hidden web shell and targeted misconfigured internal Hadoop/HiveServer2 services that accepted default authentication. While the AI agent executed repetitive tasks like crawling through personnel records and scanning for kernel vulnerabilities, the operator manually developed targeted exploit scripts and password lists. Although evidence suggests that the infrastructure was linked to prior malicious activities, there was no confirmation of data exfiltration at the time of the discovery.

In light of this incident, administrators are advised to audit their internal Hadoop clusters to ensure that authentication settings are not left as "NONE" and to restrict the installation of unauthorized user-defined functions. Security teams should also monitor web servers for any suspicious connections to internal database ports and search for hidden PHP files within web directories. Additionally, it is crucial to patch systems against recent Linux kernel vulnerabilities and legacy flaws to prevent the privilege escalation paths that automated tools can exploit.