// threat intel

Silent Lynx AI Agent Attack: DeepSeek-Powered Espionage in Central Asia

Silent Lynx AI agent attack: Hermes agent on DeepSeek models used for Tajikistan espionage, exposed server infographic

The first documented Silent Lynx AI agent attack came to light through a leak: CloudSEK found an exposed group server holding more than a dozen gigabytes of stolen government documents, hacking tools, and logs of an AI agent automating an espionage operation against Central Asian governments. The agent, an open-source framework running DeepSeek models, needed no novel skills to support the intrusion.

Key takeaways

  • CloudSEK found an exposed Silent Lynx server with 12+ GB of stolen government documents, tools, and AI-agent automation logs.
  • The operator used the open-source Hermes agent on DeepSeek models for recon, data collection, and anti-forensic cleanup, with some destructive commands pre-approved.
  • One railway employee's reused credential gave access to Tajikistan's interagency document platform; automated tools pulled thousands of Defense Ministry documents dated 2026.
  • No novel exploit was used: misconfiguration, credential reuse, and flat interagency trust. Kazakhstan ties are medium confidence; no Kazakh government link.

What did the exposed Silent Lynx server reveal?

CloudSEK, a Singapore-based cybersecurity firm, discovered an unprotected Silent Lynx server containing stolen government documents, hacking tools, and records showing how an attacker used an AI agent to automate parts of the operation. Recorded Future News reported the findings on Wednesday.

The haul documents espionage against Central Asian governments and intrusion into critical infrastructure, chiefly Tajikistan’s national railway. The campaign is visible only because the operators’ own server leaked.

How did the Silent Lynx AI agent attack use Hermes?

The operator ran Hermes, an open-source AI agent configured as an offensive cybersecurity expert, primarily on models from China’s DeepSeek. Unlike a chatbot, the agent could execute commands and run software. It handled reconnaissance, data collection, and erasing traces of activity on compromised systems. Some potentially destructive commands were configured to run without additional approval.

This is a standard agent architecture: a model in a tool-calling loop with shell, HTTP, and file access, plus a system prompt. The offensive capability came from the permission model, not model sophistication. The operator issued instructions in Russian through a Telegram bot, controlling the operation from a phone.

For incident responders, hunt shell history, temp scripts, and egress to model API endpoints. Trace erasure means scripted deletion of logs and history files: bash_history gaps and journald rotation anomalies.

One railway credential became thousands of Defense Ministry documents

The kill chain required zero novel exploits. Silent Lynx held extensive access to Tajikistan’s national railway network from at least mid-August through early October; the source does not state the year. It moved across several servers, extracted credential databases, and stole internal documents and operational reports.

Credentials of a single railway employee then worked on Tajikistan’s national electronic document management platform, the interagency exchange system ministries use for official correspondence. The group built automated tools to bulk-download files, pulling thousands of Defense Ministry documents dated 2026, plus correspondence tied to the National Guard, the State Committee for National Security, and the State Investment Committee. The server also held information about more than 100 government organizations and their employees.

The root cause is flat trust in interagency exchange: one account touching both the railway and the platform could read documents addressed to other agencies. CloudSEK also found evidence of compromise at a server supporting Uzbekistan’s Smart Forestry initiative, and malicious files disguised as official Turkmenistan documents (a railway ticket, a regional government report), created and tested in October, likely preparation for a new phishing wave. Whether that campaign succeeded is unknown.

Who is Silent Lynx?

Silent Lynx is also tracked as Cavalry Werewolf and YoroTrooper and has been active since at least 2022, previously targeting government and diplomatic organizations in Kyrgyzstan, Turkmenistan, Tajikistan, Armenia, Belarus, Russia, and Europe. In October 2025 it ran a months-long phishing campaign against Russian government agencies and industrial firms using emails disguised as Kyrgyz government correspondence.

CloudSEK assesses with medium confidence that the group has ties to Kazakhstan, citing Russian-language communications and test connections from a residential internet provider in the country. No connection to the Kazakh government was established.

Why this attacker profile should worry defenders

A human directed the operation from a phone; the agent lowered the labor cost of recon, collection, and cleanup. The access came from misconfiguration, credential reuse, and missing segmentation between an operational network and an interagency platform. Concrete fixes from this chain:

  • Enforce MFA and conditional access on document-management platforms; a single reused password should not open interagency mail.
  • Tune bulk-download detection in EDMS logs: alert on high document-fetch rates per session and sequential ID enumeration.
  • Segment operational networks (railway, utilities) from interagency document exchange; one foothold should not traverse both.
  • Deny-list destructive shell commands and require human approval for them in any AI agent configuration.

Silent Lynx pre-authorized destructive command execution in an agent config, the same over-permissioning we see in corporate agent rollouts. If your internal agents have shell access with auto-approve, you have built the attacker’s toolkit with a different prompt. See AI agent least-privilege design, LLM-orchestrated C2 in a real intrusion, and agentic AI threat modeling.

FAQ

What AI agent did Silent Lynx use?

Hermes, an open-source agent on DeepSeek models, controlled in Russian via a Telegram bot.

Did the Silent Lynx AI agent attack run autonomously?

Not fully. A human operator directed it remotely, but the agent executed certain potentially destructive commands without additional approval and automated recon, collection, and trace erasure.

How long did Silent Lynx have access to Tajikistan's railway network?

At least mid-August through early October; the source does not state the year.

Is Silent Lynx linked to the Kazakh government?

No. CloudSEK assesses with medium confidence that the group has ties to Kazakhstan, but no connection to the Kazakh government was established.

If your organization runs AI agents with tool access, or operates an interagency document platform, we test both: see our AI agent penetration testing to find what your agents can really do before someone else does.


// get started

Work with AgentOffense

Tell us about your target and goals. We’ll reply with scope and a fixed-price quote — usually within one business day.

./request_engagement