AiStaffo

Kazakh-linked hackers used AI agent to steal 2,000 Tajik files

Stock photo: server rack cables
Photo: Brett Sayles / Pexels

Singapore-based cybersecurity firm CloudSEK said in a report published on 11 November 2026 that a hacking group it links to Kazakhstan used Hermes, an open-source AI agent framework, to steal more than 2,000 national security-related documents from Tajikistan. According to CloudSEK, the group, called Silent Lynx, used the agent to plan attacks, write programs, run them and produce reports. The operation came to light after the operators left a server exposed to the internet. Other researchers have tied Silent Lynx to Kazakhstan, but the report does not show that the Kazakh government was involved. Neither government replied to requests for comment. For business owners, the lesson is practical: an AI agent with broad access and no supervision can act at machine speed, so access rights, logs and server security matter as much as the tool itself.

In short

  • CloudSEK says Silent Lynx used the open-source Hermes agent to plan, script and run an intrusion that took more than 2,000 Tajik documents.
  • The operation was found after an exposed server was left open to the internet.
  • Attribution to Kazakhstan is a medium-confidence assessment, and no link to the Kazakh government has been shown.
  • Any agent with broad account access can act at machine speed, so access limits and logs matter most.
  • Review every account and key an agent can use before adding more automation.

What happened

According to CloudSEK, a hacking group known as Silent Lynx used Hermes to scout targets from around mid-August 2026. A few weeks later, the group reached Tajikistan's electronic document system through a railway employee's account. Reuters, which carried the report, said the stolen material included national security-related files. The documents were tied to Tajikistan's Defence Ministry, according to The Times of Central Asia.

CloudSEK said it found the activity after the operators left a server open to the internet without authentication. Its own write-up, titled Off the Rails, describes an open directory holding the group's tools, stolen files and the agent's working files, covering activity through October 2026. The firm has said the exposed data included instructions the operator gave the agent about how it should behave.

Who is behind it

CloudSEK assessed with medium confidence that Silent Lynx has ties to Kazakhstan. The firm pointed to Russian-language communications and test connections from a residential internet provider in Kazakhstan, according to The Record. Other researchers have made similar links. The Record also notes that these findings do not establish any connection to the Kazakh government. Silent Lynx has been active since at least 2022, and other researchers track it under other names, including YoroTrooper and Cavalry Werewolf.

Reuters reported that efforts to reach the group were unsuccessful, and that Kazakh and Tajik officials did not respond to requests for comment.

Why it matters for businesses that automate work

The report is about espionage, not office software. Still, it shows the pattern that matters for any business that runs automation. CloudSEK said the operator did not type commands by hand. The operator gave the agent goals, and the agent wrote the scripts, ran them, kept notes and drafted reports. Reuters quoted CloudSEK describing this as the attacker telling Hermes what to do while Hermes planned, ran and reported on the attack.

The same features that make an agent useful in a business, such as memory, scripts and scheduled runs, made this operation faster. The agent also reportedly wrote a program to download files, sort them and split them into chunks. That is the kind of bulk file handling a finance or operations team may ask an agent to do for invoices or reports.

What changes in practice for a business that automates work

The first question is not which AI model to use. It is what the agent can reach. The Tajik intrusion began with one employee account, according to CloudSEK's account as reported by Times of Central Asia. A single login with wide rights was enough to reach far beyond the railway's own scope. Businesses of any size can take three steps from this.

  1. List every account, API key and shared folder an agent can open, and remove rights the task does not need.
  2. Keep agent credentials out of plain text files and shell commands, and rotate them on a set schedule.
  3. Keep a readable log of every action an agent takes, and have a named person review it on a set schedule.

Small firms can do much of this with the admin settings of the tools they already use. Larger groups usually need a security team to set the rules and review exceptions.

AreaWhat to checkWho checks it
AccessWhich accounts and folders the agent can openOwner or IT lead
SecretsWhere passwords and keys are storedIT or finance system admin
ServersWhether any agent host is reachable from the internetIT or hosting provider
LogsWhether every action is recorded and reviewedNamed manager

Where automation is not the answer

An AI agent is a poor fit for any task where a wrong action is costly and hard to reverse, such as moving money or changing records that others rely on, unless a person approves each step. Agents can make the wrong call with confidence. This case also shows that an agent can run for weeks without anyone noticing, which is a risk for any business that leaves an agent running on a server.

What to watch next

Three things are worth tracking. The first is whether any government confirms the intrusion or responds to the report. As of the reports published on 11 November 2026, neither Kazakh nor Tajik officials had responded. The second is whether CloudSEK or other researchers publish more detail on the Hermes and DeepSeek setup, which some outlets say the server records show. The third is whether security vendors and regulators issue guidance on agent permissions. Business owners should watch for that guidance because it will likely shape what their IT providers offer.

How AiStaffo would automate this

AiStaffo would start by mapping every account, folder and key that an automated worker can reach, then cut those rights to the minimum each task needs. The AI workers it builds would run on a server that is not reachable from the open internet, with credentials held in a managed secrets store rather than in scripts. Each action would be written to a log that a named staff member reviews on a set schedule, and any payment, record change or outside send would wait for a person to approve it. The owner still decides which tasks are automated and who holds approval rights. To check how this would apply to your own systems, book a free automation audit.

Questions people ask

What is the Hermes AI agent?
Hermes Agent is an open-source AI agent framework made by Nous Research. It is released under the MIT licence and runs on a user's own server or computer. It can work with many AI model providers, which is why it can be set up in many different ways.
Who is Silent Lynx?
Silent Lynx is the name Seqrite gave to a cyber espionage group in January 2025, according to The Times of Central Asia. Other researchers track it as YoroTrooper and Cavalry Werewolf. The group has targeted government bodies in Central Asia and nearby countries.
Did the Kazakh government order the attack?
No source reviewed shows that. CloudSEK assessed a Kazakhstan link with medium confidence, and The Record reports that the findings do not establish a connection to the Kazakh government. Neither government had responded to requests for comment at the time of the reports.
Does this mean businesses should stop using AI agents?
The report does not say that. It shows that an agent with broad access and weak supervision can do a lot of harm quickly. Businesses can still use agents for routine work if they limit access, protect credentials and review logs.
What should a business check first?
Start with a list of every account, folder and API key that any automation can use. Remove rights that a task does not need and move secrets out of plain text files. Confirm that no agent server can be reached from the open internet.
Where can I read the original report?
CloudSEK published its findings as Off the Rails: Inside an Exposed Silent Lynx Server on 11 November 2026. Reuters reported the same story that day, and the links in the sources list lead to the coverage.

Book a free automation audit

Thirty minutes. We look at one process you run every week and tell you exactly what an AI worker would take off your desk, and what it would not.

We use these details only to reply to this request and prepare your audit, and keep them for up to two years. We send no marketing unless you tick the box above. Privacy policy

ai securityai agentsai governance