On 11 November 2026, CloudSEK's threat research team published an analysis of an exposed server it attributes, with medium-high confidence, to the espionage group known as Silent Lynx. The open directory held stolen government documents, hacking tools and the working files of Hermes, an open-source autonomous AI agent the operator had configured as an offensive assistant. According to CloudSEK, the operator held domain-level access to Tajikistan's national railway network from mid-August to at least 10 October 2026, reused one railway employee's account to reach the national electronic document management platform, and collected more than 2,000 Ministry of Defence-related documents. The agent analysed reconnaissance, tracked the intrusions, wrote itself a handoff file so it could keep collecting without the operator, and ran clean-up on compromised systems. Its settings had been loosened so destructive commands ran without confirmation. No victim has commented publicly, and the findings rest on CloudSEK's analysis.
Key takeaways
- CloudSEK reported on 11 November 2026 that an exposed Silent Lynx server showed months of intrusions against Central Asian government bodies, with activity recorded from April to October 2026.
- Stolen domain credentials did the heavy lifting: CloudSEK found plaintext passwords, NTLM hashes and directory dumps that included the Administrator and krbtgt accounts, plus default passwords for upstream ISP equipment.
- A Hermes AI agent, running DeepSeek and Gemini models and tasked over Telegram, automated reconnaissance analysis, document collection and anti-forensic clean-up. The operator had switched off confirmation prompts and the tool-loop stop.
- CloudSEK says the agent's memory told it to pull and use secrets inside a single shell command so that Hermes' own token and key redaction would not apply. No victim organisation has publicly confirmed the intrusion.
- The identity lesson: once an attacker holds domain and service credentials, an AI agent multiplies how much they can do with them. Defenders need to detect credential reuse at machine speed, not at analyst speed.
At a glance
| Organisations | Tajikistan's national railway network and national electronic document management (EDO) platform; a Smart Forestry GeoServer host in Uzbekistan; phishing lures built against Turkmenistan targets (not a confirmed victim), according to CloudSEK |
|---|---|
| When | Activity recorded April to October 2026; railway access mid-August to at least 10 October 2026; CloudSEK report published 11 November 2026 |
| Attacker | Silent Lynx (also tracked as UAT-6275 by Cisco Talos and TAG-193 by Recorded Future), per CloudSEK at medium-high confidence; a Kazakhstan nexus is assessed at medium confidence; motive assessed as intelligence gathering |
| Entry point | Not fully disclosed. CloudSEK cites the group's known LNK-to-PowerShell delivery pattern; inside the railway network the operator used stolen domain credentials and one employee's account to reach the EDO platform |
| Identities abused | Domain accounts including Administrator and krbtgt (NTDS dumps), plaintext passwords and NTLM hashes, a railway employee's standard-user account, default credentials on ISP network equipment, tunnelling keys and SMB credential stores |
| Impact | More than 2,000 Ministry of Defence-related documents, smaller sets tied to the National Guard and State Committee for National Security, hundreds of State Investment Committee letters, and a directory export of more than 100 government organisations, according to CloudSEK |
| Category | NHI; Agentic AI and AI agents. Incident class: confirmed AI-agent breach (attacker-operated agent automated collection and clean-up on compromised systems; findings from researcher analysis, not victim statements) |
What happened
Silent Lynx is an espionage group that earlier research has linked to operations against governments in Central Asia. CloudSEK's Threat Research and Information Analytics Division found an open directory on a server it ties to the group with high confidence, based on implant code markers and build paths that match published samples. The Record, reporting on 12 November, says the server held more than a dozen gigabytes of data, including stolen government documents, hacking tools and records of how an AI agent automated parts of the operation. CloudSEK links the directory to Silent Lynx at medium-high confidence, citing malware markers, the delivery pattern and a victim already named in public reporting.
The directory, CloudSEK says, recorded activity from April to October 2026. The operator ran one tunnel into the Tajik railway network in April and May 2026 and a second from mid-August to at least 10 October 2026, with domain-level access. The haul included plaintext passwords, NTLM hashes and NTDS-format dumps covering the Administrator and krbtgt accounts. From one railway employee's standard-user account, the operator built a login chain into the national EDO platform used by ministries and state bodies, and collected from it throughout 2026. CloudSEK also describes access to a host behind Uzbekistan's Smart Forestry GeoServer through a webshell and PostgreSQL, and default-password credentials for upstream ISP equipment. Lures impersonating Turkmenistan Railways and the Ahal regional administration were built and tested in October 2026, which CloudSEK reads as a planned campaign; it does not say those lures succeeded.
The agent is what puts this on an AI agent list. Hermes is an open-source autonomous agent framework, installed on the same server and renamed by the operator. CloudSEK says it had persistent memory, a task board for sub-agents, scheduled jobs and a library of offensive security skills, and that it was given a red-team persona. Its safety settings had been loosened: destructive commands ran without confirmation, the tool-loop hard stop was disabled and unsafe JavaScript evaluation was allowed. DeepSeek V4.1 Flash, reached through OpenRouter, was the primary model, with Gemini 2.5 Flash used for sub-agents. The operator tasked it over Telegram, which The Record notes meant instructions could be sent from a phone.
CloudSEK says the agent analysed reconnaissance output, kept track of the operator's own intrusions, summarised access to the railway's upstream ISP, and wrote a handoff file so it could continue EDO collection without the operator. Its memory file told it to extract and use secrets within a single shell command so that Hermes' token and key redaction would not apply. In early October 2026 it ran clean-up: deleting messages, rebuilding indexes, shredding shell histories, clearing scheduled jobs and zeroing login records. It missed the exposed directory, which is how the operation came to light. The agent's notes also mention Armenian government networks filtering the operator's traffic and an October 2026 attempt against Pakistani targets where a web application firewall repeatedly banned the operator's address. On attribution, The Record says CloudSEK "assessed with medium confidence" a Kazakhstan link, and that earlier research does not establish any link to the Kazakh government. None of the named organisations has commented publicly in the sources we read.
Timeline
| Date | Event |
|---|---|
| April 2026 | First activity recorded in the exposed directory; a tunnel into the Tajik railway network runs through April and May. |
| May 2026 | Group activity against Uzbekistan's Smart Forestry, as described in October 2026 vendor research cited by CloudSEK. |
| August 2026 | A second tunnel gives domain-level access to the railway network from mid-August. |
| October 2026 | Turkmenistan-themed lures are built and tested; the agent runs anti-forensic clean-up in early October; railway access is seen until at least 10 October. |
| 11 November 2026 | CloudSEK publishes its analysis of the exposed server (first public disclosure). |
| 12 November 2026 | The Record reports the findings and CloudSEK's medium-confidence Kazakhstan assessment. |
How it happened: the identity attack path
- Foothold. CloudSEK does not set out the exact first step in the railway network, but cites the group's established pattern of LNK files that launch PowerShell as one of its attribution markers.
- Domain credentials harvested. The operator gathered plaintext passwords, NTLM hashes and NTDS dumps including the Administrator and krbtgt accounts, giving domain-level control and the ability to act as almost any account.
- One user account reused across systems. A railway employee's standard-user account became a login chain into the national EDO platform, a separate system used across government, where documents were collected over months.
- Network equipment and adjacent hosts. Default passwords on upstream ISP routers and switches, a webshell and database access on the Uzbek host, and stored tunnelling keys and SMB credential databases widened the operator's reach.
- Agent automation. A Hermes agent with loosened safeguards worked the access: summarising reconnaissance, continuing collection from a handoff file, and handling secrets in a way designed to slip past its own redaction.
- Clean-up and exposure. The agent wiped histories, scheduled jobs and login records on compromised systems, but the operator's own server stayed open to the internet, exposing the tools and data.
Impact
- Reported by CloudSEK: more than 2,000 unique Ministry of Defence-related documents dated through 2026, covering flight clearances, personnel transfers, attaché correspondence, defence procurement and passport and national ID scans; smaller sets tagged to the National Guard and State Committee for National Security; hundreds of State Investment Committee letters; a directory export of more than 100 government organisations; and railway operational reports and internal correspondence.
- Not confirmed by victims: none of the organisations named has publicly confirmed the intrusion or its scope in the sources we read. The figures are CloudSEK's, drawn from the operator's own files.
- Attempted or planned: lures against Turkmenistan targets were built and tested, and the agent's notes describe blocked attempts against Armenian and Pakistani networks. CloudSEK does not report success there.
- Wider: the case shows an espionage operator handing routine collection and anti-forensics to an AI agent, which shortens the time between credential theft and data loss and removes traces defenders rely on.
What this means for NHI and AI agent security
Two kinds of non-human identity sit at the centre of this case. On the victim side, domain service accounts such as krbtgt, stored credentials, tunnelling keys and default passwords on network devices gave the operator lasting, quiet access. A compromised krbtgt account in particular lets an attacker mint Kerberos tickets until it is reset twice, so any organisation that suspects an NTDS dump should treat it as a full domain compromise. On the attacker side, an autonomous agent with its own model API access, memory and scheduled jobs acted as a tireless operator, keeping collection going while the human was away.
The agent detail matters for defenders building their own agents too. The operator disabled confirmation prompts and the tool-loop stop, and taught the agent to dodge the framework's secret redaction. Those are exactly the controls that organisations rely on when they deploy agents with real credentials. Safeguards that live inside the agent can be switched off by whoever controls it, so policy enforcement, credential brokering and logging need to sit outside the agent. Our Agent Controls and Monitoring Guide and AI-Powered Cyber Attacks Guide cover both sides of this. Similar agent-driven credential abuse appears in the Chainalysis breach 2026 and the ARTEX agentic secret harvesting campaign.
Finally, the exposed directory is a reminder that attacker infrastructure leaks too. Threat intelligence teams that watch for open directories can sometimes see an operation from the inside, as CloudSEK did here.
Recommendations
- Treat an NTDS dump as total domain compromise. Reset the krbtgt account twice with a safe interval, rotate all privileged and service account passwords, and rebuild trust from clean hosts. See our Active Directory Domain Compromise Response Playbook and krbtgt account guide.
- Remove default credentials from network equipment. Inventory routers, switches and appliances, including those run by upstream providers on your behalf, and require unique credentials with central authentication.
- Stop one account opening many systems. Separate identities and access for platforms such as shared government document systems, enforce strong multi-factor authentication there, and alert when a standard user suddenly pulls large volumes of documents.
- Detect machine-speed collection and log wiping. Alert on bulk downloads, cleared shell histories, deleted scheduled jobs and zeroed login records, and ship logs off host so local clean-up cannot erase them. See our Agent Threats and Attacks Guide.
- Put agent safeguards outside the agent. For your own agents, enforce approvals, tool limits and secret handling in a gateway or broker the agent cannot reconfigure, and keep credentials out of the agent's shell. See our Agent Sandboxing vs Tool Allowlisting comparison.
- Use the published indicators. CloudSEK lists the server address, domain and implant hashes; search for them in network and endpoint telemetry, especially in Central Asian government and transport networks.
Frequently asked questions
What is Silent Lynx?
Silent Lynx is an espionage group that researchers have linked to operations against government bodies in Central Asia and neighbouring countries. It is also tracked as UAT-6275 by Cisco Talos and TAG-193 by Recorded Future, and CyberDelegate reports the names Cavalry Werewolf and YoroTrooper. CloudSEK assesses a Kazakhstan nexus at medium confidence, but earlier research has not established any link to the Kazakh government.
How did Silent Lynx use the Hermes AI agent?
According to CloudSEK, the operator ran Hermes, an open-source autonomous agent framework, on its own server with DeepSeek and Gemini models and tasked it through Telegram. The agent analysed reconnaissance, tracked intrusions, continued document collection from a handoff file and wiped traces from compromised systems. Its confirmation prompts and tool-loop stop had been turned off, and it was told to handle secrets in a way that avoided Hermes' redaction.
Have the affected governments confirmed the breach?
Not in the sources we read. The findings come from CloudSEK's analysis of the operator's exposed server, reported by The Record and others. No Tajik, Uzbek or Turkmen body has publicly confirmed the intrusion or the scope of the documents taken, so the figures should be read as CloudSEK's.
Related NHI Mgmt Group resources
Chainalysis breach 2026 · ARTEX agentic secret harvesting 2026 · AI-Powered Cyber Attacks Guide · Agent Controls and Monitoring Guide · Active Directory Domain Compromise Response Playbook
How NHI Mgmt Group can help
This intrusion turned on service accounts, stored credentials and default device passwords, and an AI agent made the most of them at machine speed. NHI Mgmt Group helps security teams find and govern those machine identities, plan krbtgt and service account recovery, and design controls for AI agents that hold up even when someone tries to switch them off. See our NHI and AI agent security training.
References
- CloudSEK: Off the Rails: Inside an Exposed Silent Lynx Server (11 November 2026)
- The Record: Kazakhstan-linked hackers used AI agent to spy on Central Asian governments (12 November 2026)
- CyberDelegate: Silent Lynx espionage campaign exploits Central Asian government networks with AI-driven tools (13 November 2026)