A hacking group linked by researchers to Kazakhstan used an autonomous AI framework during a cyber-espionage operation that stole more than 2,000 national-security-related documents from Tajikistan, according to a Reuters report published November 11.
The reporting is based on research from Singapore-based cybersecurity company CloudSek. It does not establish that the governments involved authorised or carried out the operation, and officials from Kazakhstan and Tajikistan had not immediately responded to requests for comment.
What the researchers found
CloudSek said it identified the activity after the attackers inadvertently exposed a server to the internet. The group, which CloudSek called “Silent Lynx,” used Hermes, an open-source framework for autonomous AI agents.
According to the report, the agent handled reconnaissance beginning around mid-August and later helped break into Tajikistan’s electronic document system. The recovered data indicated that Hermes planned attacks, wrote rules, ran them, kept notes and prepared reports. It also wrote a custom program to download, organise and split files into chunks.
CloudSek linked the activity to a group researchers associate with Kazakhstan based on evidence including Russian-language activity, Kazakh business hours and access through a residential Kazakh internet connection. Those are indicators, not conclusive public proof of the operators’ identity.
Why this matters
The significance is not that AI made a cyberattack possible. Attackers have automated reconnaissance and data theft for years. The shift is that an agent can coordinate multiple steps, adapt instructions and maintain operational notes with less hands-on direction from the operator.
That can lower the effort required to run long, targeted intrusions and make the activity harder to distinguish from ordinary administrative automation. It also means defenders must examine not only malware signatures, but unusual sequences of actions, access patterns and tool use.
Defensive implications
Organizations should assume that exposed management interfaces and document systems will be tested by automated systems. Strong identity controls, network restrictions, rapid patching, detailed audit logs and detection of abnormal bulk access remain more important than any single AI-specific product.
They should also protect the AI tools used internally. If an agent can browse systems, execute code or move files, its permissions should be task-scoped and independently monitored. An agent that inherits a user’s full access can turn a compromised session into a much larger incident.
The case is an early public example, not evidence that every threat actor now operates autonomous agents. But it is a warning that defensive playbooks built around slow, manual attacker behaviour are becoming less reliable.


