Kazakhstan-Linked Hackers Use AI Framework Hermes to Autonomously Steal 2,000 Files from Tajikistan

On November 11, 2026, CloudSek reported that the Kazakhstan-linked hacking group Silent Lynx has since mid-August used the open-source AI agent framework Hermes to attack Tajikistan and steal 2,000 classified national-security documents.

Reconstructing the Facts

According to CloudSek, the attack began with a reconnaissance phase in mid-August and, several weeks later, breached Tajikistan’s electronic document management system. The Hermes framework autonomously planned paths, wrote scripts, packaged files into chunks, and exfiltrated them without continuous human control. The attackers were discovered after one server was accidentally exposed. From the exposed server, CloudSek recovered the hackers’ instructions to the AI agent, which directed the agent to act as an “elite offensive security expert and red team member.”

Mechanism Breakdown

CloudSek described that the attackers gave Hermes only high-level tasks rather than entering commands one by one. Hermes then independently carried out reconnaissance, rule writing, execution, logging, and report generation. The entire process shows that the AI agent was configured as a standalone operating unit, reducing the need for human intervention. Silent Lynx members used Russian, followed Kazakhstan working hours, and connected to servers through residential networks.

Industry Impact

This case shows that open-source AI agent frameworks have entered nation-state espionage toolchains. Hermes’s autonomous capabilities allow attackers to outsource complex tasks and reduce operational traces. Vulnerabilities in electronic document management systems were quickly exploited by the AI agent, exposing the limitations of traditional manual monitoring when facing autonomous agents. Other research organizations had previously linked Silent Lynx to Kazakhstan, and this incident further confirms its scope of activity.

Strategic Assessment

[Analysis] The Hermes case shows that the boundaries of AI agents’ “instruction-following” can be exploited in reverse: attackers use role-setting instructions to guide agents into performing offensive tasks rather than relying on continuous human oversight. This contrasts with traditional APT operations, in which manual monitoring is greatly reduced. In the future, defenders will need to build new detection mechanisms targeting AI agent behavior patterns rather than relying solely on existing log audits. Officials from Kazakhstan and Tajikistan have not yet responded to the reports.