AI Agents Automate Post-Exploitation Task

A recent discovery by threat intelligence firm Hunt.io and security researcher Bob Diachenko has shed light on how modern offensive operations are evolving. By analyzing exposed web directories, researchers identified evidence that attackers utilized an open-source AI agent to facilitate post-exploitation activities against the Thailand Ministry of Finance.

Infrastructure and Exposure

Between July 9 and July 13, investigators identified three improperly secured server directories hosted in Hong Kong. These repositories housed 585 files, totaling roughly 470 MB of data. The contents included custom exploit code, web shells, and specialized payloads, alongside logs that directly linked the infrastructure to the ministry’s internal systems, including Hadoop environments and mail servers.

Researchers utilized unique identifiers, specifically JA4X fingerprints derived from TLS certificates, to map the breadth of the attacker-controlled network. This led to the discovery of additional servers in Malaysia and Hong Kong. While artifacts indicated clear targeting of internal services, the ministry itself has not confirmed a successful compromise of its systems.

The Role of Autonomous Agents

The most notable finding was the use of Hermes, an open-source AI agent designed to maintain persistent task memory. The attackers explicitly configured the software to operate in YOLO mode. This setting is specifically designed to bypass human-in-the-loop requirements, allowing the agent to execute potentially sensitive commands without repeated operator verification.

In addition to the common name, all these certificates share a JA4X fingerprint, a hash derived from the structure of the certificate itself rather than its contents.

— Hunt.io report

Logs recovered from the exposed directories show the agent performing complex post-exploitation tasks, such as enumerating SUID and SGID binaries, inspecting containers, and scanning for kernel-level vulnerabilities. The agent was also directed to execute custom versions of LinPEAS to gather system information. While the AI successfully cataloged sensitive documents, there is no current evidence indicating that these files were exfiltrated.

• 585 files were discovered in the exposed server directories.
• 470 MB of data was recovered from the compromised infrastructure.
• 14% is the rate at which security teams currently alert on successful attacks.

Implications for Security Operations

This incident reflects a broader trend of integrating automation into the adversary lifecycle. The use of agents like Hermes—or the recently noted JadePuffer ransomware operation—indicates that threat actors are shifting toward high-speed, autonomous execution for routine tasks like lateral movement and reconnaissance. As these tools become more accessible, the window of time that security teams have to detect and block a breach is narrowing.

For organizations, this shift underscores the necessity of proactive, automated testing. Because these agents can move through an environment with minimal manual input, static defenses are increasingly insufficient. The ability to verify the efficacy of SIEM and EDR rules through continuous breach simulation may become a standard requirement for maintaining network visibility against autonomous threats.