Researchers find hackers running a Central Asian spying campaign through an AI agent

Security researchers at CloudSEK say they found an open, unsecured server that let them watch a state-linked hacking group direct an espionage campaign across five Central Asian countries largely through an autonomous artificial intelligence agent, rather than through hands-on commands from a human operator.

In a report titled "Off the Rails: Inside an Exposed Silent Lynx Server," published Wednesday, the firm attributes the activity with high confidence to Silent Lynx, a Central Asia-focused espionage group also tracked by other security firms under the names UAT-6275 and TAG-193. CloudSEK says it assesses with medium confidence that the operator has ties to Kazakhstan, based on the operation running in Russian and the operator's own test traffic coming from a residential Kazakh connection — consistent with existing vendor assessments of the group.

An AI agent doing the work

According to the report, the exposed server was running Hermes, an open-source autonomous agent framework, installed under the custom name "Strix" and configured with persistent memory, a running task list and scheduled jobs. The primary model driving the agent's decisions was DeepSeek's V4.1 Flash, with Google's Gemini 2.5 Flash handling delegated sub-tasks.

What was taken

The bulk of the stolen material came from Tajikistan, where the report says the operator reached the national electronic document-management platform and government bodies including the Ministry of Defence, the National Guard and the State Committee for National Security. CloudSEK says it recovered more than 2,000 unique documents dated through 2026, most of them correspondence to or through the defense ministry. The group also compromised Tajikistan's national railway network, the report says. Elsewhere, the operator gained host-level command execution on Uzbekistan's Smart Forestry mapping platform, and had built and tested phishing lures impersonating Turkmenistan Railways as recently as October. The researchers also found an implant built for a target in Kyrgyzstan and logged a separate October intrusion attempt against Pakistan's Ministry of Foreign Affairs.

CloudSEK's report frames the find as an early, concrete example of a state-linked group using a general-purpose AI agent, rather than custom-built malware alone, to handle reconnaissance, data collection and even cleanup of its own log files — though the researchers note the cleanup effort "missed the directory that was exposed," which is how the campaign came to light at all. None of the governments named in the report had issued a public response as of Wednesday.