DN DreamNoos Stories for Curious Minds. Independent Digital Magazine
Science & Technology Cybersecurity

Hackers Are Using AI Agents to Steal Government Documents, New Investigation Finds

A new CloudSEK investigation reveals how a suspected Central Asian espionage group used autonomous AI agents to target government systems and steal sensitive documents.

Dark server room with cybersecurity monitoring equipment
Photo by Tyler on Unsplash

Cyber espionage is entering a new phase. Hackers are no longer relying exclusively on manually executed commands and conventional malware to carry out complex operations. Increasingly capable AI agents are beginning to take on parts of the work themselves " from reconnaissance and data collection to organizing stolen files.

By Alper KALE
November 11, 2026

A new investigation published by cybersecurity firm CloudSEK on November 11 has provided a detailed look at how this approach may be used in operations targeting government institutions and critical infrastructure across Central Asia.

The investigation centers on a suspected espionage group known as Silent Lynx. According to CloudSEK, the group’s exposed server contained evidence of intrusions involving Tajikistan’s government document systems, railway networks and an Uzbek government-operated platform.

Reuters separately reported that the group used AI to obtain more than 2,000 national security-related documents from Tajikistan.

The findings raise an important question for governments and security teams: what happens when an attacker can delegate significant parts of a cyber operation to an autonomous AI system?

An exposed server revealed the operation

CloudSEK said it discovered an internet-accessible server containing the group’s tools, operational records, collected files and conversations with an AI agent.

The exposed directory offered researchers an unusual view into the operation. Rather than relying only on malware samples or traces left on victim systems, investigators could examine the operator’s own working environment.

According to the report, the server contained records spanning several months of activity in 2026. CloudSEK linked the infrastructure and tools to the Silent Lynx cluster through similarities with previously documented malware and operational techniques.

The researchers assessed with high confidence that the exposed directory belonged to a Silent Lynx operator. They also identified evidence suggesting a Kazakhstan connection, although geographic attribution in cyber investigations requires careful interpretation.

A language pattern, working-hour schedule or network connection can support an assessment, but none should automatically be treated as proof of an individual’s identity or a government’s involvement.

The AI agent was given more than a supporting role

One of the most significant details in the investigation is the way the attackers used an autonomous agent framework called Hermes.

The framework was not simply being used to answer questions or generate snippets of code. According to CloudSEK, the operator delegated multiple operational tasks to the agent, including reconnaissance, information collection, writing supporting scripts and organizing the results.

The report describes a self-hosted setup in which DeepSeek models were used through OpenRouter for much of the recorded activity. The agent also had access to tools that allowed it to interact with the operator’s environment.

That distinction matters.

Traditional automation generally follows predefined instructions. An autonomous agent can be given a broader objective, break it into smaller tasks and use available tools to work toward that objective.

In this case, CloudSEK’s evidence suggests that the agent was integrated into an existing espionage workflow rather than operating as an isolated chatbot.

It does not mean the AI independently decided to attack a government. Human operators configured the environment, selected targets and directed the activity. The AI system became part of the execution process.

More than 2,000 government documents were collected

The most consequential exposure described by Reuters involved Tajikistan’s electronic document management platform, which is used by ministries and other state institutions to exchange official correspondence.

CloudSEK said the operator accessed the platform using credentials belonging to a railway employee. The researchers identified weaknesses in document-level authorization and the way file attachments were made available.

The collected material reportedly included more than 2,000 unique documents addressed to or routed through Tajikistan’s Ministry of Defence.

The documents covered a range of administrative and security-related subjects, including personnel transfers, military correspondence, procurement inquiries, budget material and coordination involving airspace and telecommunications.

The scale and nature of the material suggest that the operation was not limited to stealing a handful of files. It involved collecting information from systems used for government communication.

CloudSEK also described access to Tajikistan’s railway network and an intrusion involving Uzbekistan’s Smart Forestry platform.

These findings point to a broader pattern: government systems, transport networks and digital public services can become connected parts of an intelligence-gathering operation.

Why AI-assisted cyber espionage changes the risk

The use of AI in cyber operations is not entirely new. Security researchers have previously documented attempts to use language models for phishing, malware development, research and other forms of malicious activity.

What makes this investigation notable is the reported integration of an AI agent into a longer operational process.

An agent that can coordinate tasks, retain working context and interact with software tools may reduce the amount of repetitive work required from an operator. It can also help organize complex activities that would otherwise require several separate scripts and manual steps.

That does not make every AI agent a sophisticated hacker. The effectiveness of such systems still depends on access, configuration, the quality of the underlying tools and the operator’s decisions.

But the security implications are difficult to ignore.

Organizations have traditionally focused on protecting individual applications, accounts and network boundaries. They must now also consider how stolen credentials, poorly configured services and exposed administrative tools could be combined with increasingly automated workflows.

The problem is not simply that AI can write code faster. It is that AI agents can connect different stages of an operation.

The human factor remains central

Despite the technical complexity described in the report, several weaknesses were reportedly rooted in familiar security problems.

CloudSEK identified exposed infrastructure, weak access controls, credentials that provided broader access than necessary and sensitive operational material stored alongside other files.

These are not entirely new vulnerabilities. They are examples of security failures that can become more damaging when an attacker has automated tools to exploit them.

For organizations handling government records or critical infrastructure, basic controls remain essential: strong identity management, least-privilege access, secure document authorization, network monitoring and careful handling of credentials.

Security teams should also review how autonomous agents are permitted to use internal tools. An AI system with broad access and insufficient approval requirements can introduce risks even when it is deployed for legitimate purposes.

A warning about the next generation of cyber operations

The CloudSEK investigation does not establish that AI has replaced human hackers. Instead, it illustrates how human-directed cyber operations may become more automated and more difficult to investigate.

The suspected Silent Lynx operation combined conventional intrusion techniques with an AI agent capable of handling parts of the workflow. That combination is likely to attract attention from security researchers and government defenders well beyond Central Asia.

For now, the central lesson is straightforward: organizations cannot treat AI security and traditional cybersecurity as separate concerns.

As autonomous tools become more capable, protecting systems will require not only stronger defenses against malware and stolen credentials, but also tighter control over what automated agents can access, execute and change.

The next major cybersecurity challenge may not be a hacker working faster at a keyboard. It may be a human operator directing several automated systems to work together.


Author: Alper KALE
Publication: DreamNoos
Article type: News Analysis

Sources:

  1. Reuters " Kazakhstan-linked hackers used AI to steal Tajik government documents
  2. CloudSEK " Off the Rails: Inside an Exposed Silent Lynx Server

Image note: Representative image of a dark server room. Does not depict specific government systems or breached organizations.

Reader case studies

Anonymised composites from reader correspondence and editorial review — names and identifying details removed. They illustrate how layered reads apply in practice.

  1. A retiree adjusting to a recent move reported dreaming of Hackers Are Using AI Agents to Steal Government Documents, New Investigation Finds after an anniversary date approaching. On waking review, she connected the scene to burnout rather than prophecy; Islamic scholarship context helped separate ru'ya from ordinary stress imagery.

  2. An artist between commissions reported dreaming of Hackers Are Using AI Agents to Steal Government Documents, New Investigation Finds after a move to a new neighbourhood. On waking review, she realised the dream tracked grief she had postponed; classical and psychological layers pointed the same direction.

These are editorial teaching examples, not testimonials or medical case reports.

THE WORLD IS FULL OF STORIES

Stay curious. Keep discovering.

Explore DreamNoos ↗