Silent Lynx uses AI agent to raid Central Asian rail networks

A CloudSEK investigation exposes how Silent Lynx weaponised a DeepSeek-powered AI agent to breach Tajik rail and defence infrastructure.

A large control room features a multi-panel video wall displaying complex data visualizations, above rows of empty office desks each with multiple computer monitors, illuminated by natural light from floor-to-ceiling windows.

A cyber espionage group known as Silent Lynx has compromised Tajikistan's national railway network, collected more than 2,000 defence-related government documents, and breached a geospatial platform operated by the Uzbek government. The campaign, uncovered by threat-intelligence firm CloudSEK through an exposed Silent Lynx server, is notable for a reason that extends well beyond Central Asia: the attackers used a self-hosted AI agent to assist with reconnaissance, data collection and operational clean-up, an early operational-scale instance of AI being deployed as an active tool in state-linked espionage.

The exposed server held more than a dozen gigabytes of material, including malware samples, compromised credentials, stolen government records and operational logs spanning April to October 2026. The breadth of the haul points to a methodical, sustained campaign rather than an opportunistic breach.

Railway infrastructure as a strategic target

CloudSEK's researchers found that Silent Lynx maintained domain-level access to Tajikistan's national railway network from mid-August until at least 10 October 2026. Attackers extracted administrative credentials and collected internal operational reports across multiple servers. In a separate intrusion, the group used a railway employee's credentials to access Tajikistan's national electronic document-management platform. Weak document-level access controls allowed the collection of thousands of government records. More than 2,000 of those documents originated from the Ministry of Defence, covering military personnel transfers, flight-clearance requests, defence procurement correspondence, air-defence cooperation records and identity documents.

The campaign also reached Uzbekistan, where attackers compromised the government's Smart Forestry GeoServer, achieving command execution, database access and persistent foothold. Phishing lures impersonating Turkmenistan Railways and a regional government administration were also discovered, suggesting the group was preparing further operations.

AI as an espionage instrument

The most strategically significant finding is the confirmed use of a self-hosted AI agent built on the Hermes framework and powered primarily by DeepSeek models. The operator used the agent to support reconnaissance, analyse collected intelligence and perform operational clean-up, including the ability to execute certain destructive commands without additional approval. This is not AI as a marketing adjective; it is AI embedded in an adversary's kill chain.

The implications stretch across sectors. For critical-infrastructure operators, railways, energy grids, port logistics, the threat model has shifted. Previously, the labour cost of sustained espionage imposed a natural ceiling on the volume of data an attacker could meaningfully process. AI agents erode that ceiling, allowing smaller teams to collect, sort and act on intelligence at scale. For the defence procurement and transportation-logistics communities, the Silent Lynx campaign illustrates a convergence risk: the same open-weight AI models being adopted to optimise supply chains can be re-purposed as adversarial reconnaissance tools.

CloudSEK attributed the exposed server to Silent Lynx with high confidence, citing malware code markers and build-path similarities with previously documented operations. Operator test connections traced to residential internet services in Kazakhstan, including Kazakhtelecom, alongside Russian-language operational instructions, support earlier assessments of a Kazakhstan nexus, though the firm assigns that attribution medium confidence only.

The capital and regulatory read-across

For investors and policymakers, the story sits at the intersection of two accelerating dynamics. First, the proliferation of capable open-weight AI models means that a well-resourced threat actor no longer needs to build proprietary tooling, the barrier to AI-assisted espionage has collapsed in line with the barrier to AI-assisted development. Second, critical-infrastructure cybersecurity is a growth vector for defence-tech and cybersecurity vendors, with government procurement budgets in NATO-adjacent states and GCC sovereigns both expanding rapidly in response to exactly this threat profile.

The Silent Lynx campaign will add pressure on regulators in the EU and UK, where the Network and Information Systems 2 directive is already forcing rail and energy operators to lift security baselines, to accelerate implementation timelines. For cross-sector investors with exposure to both frontier AI and critical-infrastructure operators, the message is blunt: the speed at which open-weight models are being operationalised as offensive tools is outpacing the defensive upgrade cycle in most of the world's transport and logistics infrastructure.