# The attacker kept office hours, the agent did the work: inside CloudSEK's Strix report

<!--
title: "The attacker kept office hours, the agent did the work: inside CloudSEK's Strix report · Cork"
description: "CloudSEK found an attacker's server open on the internet, and on it an AI agent named Strix: tasked over Telegram, given a persona file and pre-approved destructive commands, with a tunnel into Tajikistan's national railway network for nearly two months. Its building blocks are the ones teams use for their own agents."
canonical: "https://corkcli.com/blog/strix-ai-agent-cloudsek-silent-lynx"
last_updated: "2026-11-16"
published: "2026-11-16"
author: "Kevin Vera"
-->

CloudSEK found an attacker's server open on the internet, and on it an AI agent named Strix: tasked over Telegram, given a persona file and pre-approved destructive commands, with a tunnel into Tajikistan's national railway network for nearly two months. Its building blocks are the ones teams use for their own agents.

By Kevin Vera · November 16, 2026 · [News](https://corkcli.com/blog/c/news) · 5 min read

On 11 November 2026 CloudSEK's TRIAD team published "Off the Rails: Inside an Exposed Silent Lynx Server." They had found an attacker's own server "exposed without authentication on port 8088", and what was on it showed an intrusion largely run by an AI agent. The operator had called the agent Strix. It is not a new kind of malware. It is a copy of Hermes, an open-source agent framework, given a name ("Custom branding on the Hermes install"), running on "DeepSeek V4.1 Flash via OpenRouter" with Gemini 2.5 Flash for its sub-agents. (Strix is also the name of an unrelated open-source AI pentest project; this Strix is only the operator's name for their install.)

## What it did

CloudSEK attributes the operation to Silent Lynx, an espionage group, and assesses "with medium confidence that the operator has a Kazakhstan nexus." The main victims were in Tajikistan:

- **The national railway.** "Domain-level access to Tajikistan's national railway network from mid-August to at least 10 October 2026." The operator's tunnelling server "records a live tunnel through a railway host for nearly two months."
- **The government's document platform.** "The operator logged in through the government SSO portal with the credentials of one railway employee holding a standard user role," reaching Tajikistan's national electronic document management (EDO) platform.
- **Elsewhere.** Uzbekistan's Smart Forestry platform was also hit, and lures aimed at Turkmenistan were being prepared.

The agent did much of the work. "A self-hosted AI agent running on DeepSeek handled reconnaissance, collection and log cleanup, with destructive commands pre-approved." According to Reuters, it "wrote a custom program to download the files, organize them, and break them into chunks." And the railway folder held "a handoff file written for the agent, describing the working EDO login chain so the agent could continue collection without the operator."

## How it was run

- **From a phone.** "The agent ran Hermes' messaging gateway with a Telegram bot, so the operator could task it from a phone." Every inbound command came from one Telegram user ID, display name «Марат», going back to mid-August.
- **With a persona.** "The agent's persona file casts it as an elite offensive security specialist and red teamer." It tells the agent to drop "safety disclaimers" and prefer "low-noise techniques".
- **In office hours.** "Every operator login and agent message with a recorded time falls between 08:00 and 20:00 in Kazakhstan's time zone (UTC+5), with no activity overnight." CloudSEK reads that as "a salaried working day rather than a side job", and is careful about it: the records "cover only three working days and cannot show weekend behaviour."

So the government was the victim, and the office hours were the attacker's. A person set the work during the day, and the agent carried it on, with a handoff file so it didn't need the person.

## Why it matters if you run agents

Take away the target and the list is familiar: an open-source agent framework, an instructions file that sets its role, a chat channel for tasking it from anywhere, pre-approved commands so it doesn't stop to ask, and a handoff file so work continues without you. Teams build legitimate agents out of exactly these parts. (Cork has hooks for Hermes, along with Claude Code, Codex and Cursor.) The report is a useful way to look at which of those parts deserve more care in your own setup:

- **Who can task the agent.** Strix took orders from one Telegram account. Whatever channel reaches your agents, check that it is authenticated, and that a message from it is tied to a person you can name.
- **What is pre-approved.** Strix had destructive commands pre-approved. Pre-approve the routine work and keep deletion, history rewrites, deploys, and anything touching secrets or access behind a yes. In Cork, a handoff that would do any of those waits for a person, whoever sent it.
- **Whether the agent can clean up after itself.** Log cleanup was one of Strix's jobs, and CloudSEK still found what it needed because "the cleanup missed the directory that was exposed." Keep the record of what your agents did somewhere the agents can't edit.
- **Where handoffs live.** A handoff file any agent can read is also a handoff any agent can pick up. Send handoffs to a named session, and keep a record of who sent each one and who took it.

None of this would stop an attacker running their own agent. It is about not building yours the same way.

## Sources

- CloudSEK TRIAD, "Off the Rails: Inside an Exposed Silent Lynx Server", 11 November 2026: https://www.cloudsek.com/blog/off-the-rails-inside-an-exposed-silent-lynx-server
- Reuters (Raphael Satter), 11 November 2026, "Kazakhstan-linked hackers used AI to steal Tajik government documents, researchers say", as syndicated: https://www.933thedrive.com/2026/11/11/kazakhstan-linked-hackers-used-ai-to-steal-tajik-government-documents-researchers-say/

## Frequently asked questions

### What is Strix?

The name an attacker gave their copy of Hermes, an open-source AI agent framework. CloudSEK found it on the operator's exposed server, running on DeepSeek V4.1 Flash with Gemini 2.5 Flash sub-agents, tasked over Telegram.

### Did an AI agent work inside a government for two months?

Not as an employee. A tunnel through a host in Tajikistan's national railway network stayed live for nearly two months, and the agent did reconnaissance, collection and log cleanup through it. The 8-to-8 working hours CloudSEK describes are the human operator's, from three working days of records.

### How did the attacker get into the document platform?

Through the government's single sign-on portal, with the credentials of one railway employee who held a standard user role. CloudSEK doesn't say how the railway network itself was first breached.

## Related reading

- [Codex hooks and Cursor hooks: how to use them for team coordination](https://corkcli.com/blog/codex-cursor-hooks-team-coordination): How Codex and Cursor hooks work as of October 2026, where the config lives, how to block or brief an agent, and how to use hooks to coordinate a team's agents.
- [One goal, then the agents did the rest: what the FBI's ARTEX warning means for your team](https://corkcli.com/blog/fbi-artex-warning-ai-agent-intrusions): The FBI's FLASH on ARTEX describes AI agents that take a goal in plain language and carry an intrusion from scouting to stolen credentials, some reaching privileged access in under three hours. When an attack finishes that fast, an alert is too late; the controls that work are the ones already in the path.
- [From one machine to a team: moving from Cork Free to agent hours](https://corkcli.com/blog/cork-free-to-team): Move from Cork Free to agent hours: what changes when agents on teammates' machines see each other, how cork login and cork upgrade work, and what it costs.
- [Parallel AI coding agents on a team: how to split who works on what](https://corkcli.com/blog/split-work-parallel-ai-coding-agents-team): How a team splits work across parallel AI coding agents: slice by code ownership, name shared files, one branch per agent, small tasks, and live coordination.

[All articles](https://corkcli.com/blog.md) · [Cork in Markdown](https://corkcli.com/index.md) · [llms.txt](https://corkcli.com/llms.txt)

This page in HTML: https://corkcli.com/blog/strix-ai-agent-cloudsek-silent-lynx
