Silent Lynx Hackers Used the Hermes AI Agent to Steal Tajik Files
CloudSEK says a Silent Lynx operator, linked with medium confidence to Kazakhstan, used the open-source Hermes agent framework to run reconnaissance and collection against Tajikistan's government, taking more than 2,000 documents. The agent was steered through a Telegram bot and called DeepSeek and Gemini models through OpenRouter. The operation surfaced because the attacker left the server open without authentication.
On this page
What CloudSEK found on an open server
Singapore security firm CloudSEK published a report on 11 November describing a Debian server it found exposed on the internet without authentication. The machine held more than 12 GB of material with no separation between hacking tools, stolen data, and the operator's own workspace. CloudSEK attributes the server to an operator of Silent Lynx, a long-running espionage group focused on Central Asia, with high confidence based on implant markers and build paths that match previously published samples.
Reuters, which reported the findings the same day, says the group stole more than 2,000 national security related documents from neighbouring Tajikistan. CloudSEK describes material tied to the Ministry of Defence, the National Guard, the State Committee for National Security, and hundreds of letters from the State Investment Committee, plus identity document scans for named individuals. Kazakh and Tajik officials did not immediately respond to Reuters.
How the Hermes agent ran the operation
The most striking part of the report is the division of labour. Hermes is an open-source framework for autonomous AI agents, and the operator installed it on the server and tasked it through a Telegram bot. CloudSEK told Reuters that the attacker said what to do and Hermes planned the attacks, wrote the rules, ran them, kept notes, and wrote up reports.
According to CloudSEK, the agent handled reconnaissance from about mid-August, then, after the group broke into Tajikistan's electronic document system a couple of weeks later, wrote a custom program to download the files, sort them, and split them into chunks for exfiltration. It was also used for cleaning up logs. The way in was mundane: the operator logged into a government single sign-on portal with the credentials of one state railway employee who held a standard user role, and that account reached the national document platform. Access inside the railway network ran from mid-August to at least 10 October.
The configuration shows how far the operator pushed autonomy. Hermes was allowed to run destructive commands without confirmation, its hard stop on repeated tool failures was switched off, and its persona told it to drop safety disclaimers. The exposed instructions opened by casting the agent as an elite offensive security specialist and red teamer.
Where the models actually ran
Readers interested in local AI should note one detail. The Hermes framework, its memory, scripts, and logs were self-hosted on the attacker's server, but the reasoning was not. Model calls went through OpenRouter to hosted APIs, with DeepSeek V4.1 Flash as the primary model and Gemini 2.5 Flash for sub-agents. In other words, this was not an offline attack rig; it was a self-hosted agent loop renting cheap frontier inference.
This is not the first time Hermes has turned up in an intrusion. In July, Palo Alto Networks Unit 42 described a Chinese-speaking actor who used Hermes Agent with DeepSeek as its reasoning engine to hunt for and attack exposed servers, reaching more than 460 attempted targets across autonomous and manual techniques. That operator was also exposed by its own agent, which started a file server from the home directory.
What defenders and agent builders should take from it
Two lessons stand out. First, agents lower the labour cost of espionage, but they also leave unusually rich records: notes, reports, prompts, and logs that make attribution easier once a server leaks. Second, the target side failed in an ordinary way. A single low-privilege account reached sensitive defence records because, as CloudSEK puts it, document-level access checks were weak. Per-document authorisation and monitoring of bulk downloads from one account would have mattered more than anything specific to AI.
For anyone running agent frameworks themselves, the configuration choices here are a checklist of what not to do: no confirmation on destructive commands, no loop limits, and an exposed service port. CloudSEK's report lists the server address, URLs, and file hashes for defenders who want to hunt for related activity.