An AI agent ran a nation-state espionage campaign — with its guardrails taken off
On November 11, CloudSEK's TRIAD team published Off the Rails: Inside an Exposed Silent Lynx Server. An operator from the espionage cluster known as Silent Lynx left a VPS exposed on the public internet — 12+ GB of tooling, victim data, and, most unusually, the full conversation history of the AI agent that had been running the operation. Reuters picked it up the same day.
The one line worth starting with: the agent ran with its own safeguards deliberately removed. The tool-loop hard stop was disabled. Destructive commands ran without confirmation. The operator kept the safety features of the framework, then switched them off.
What the agent actually did
CloudSEK identifies the agent as a custom-branded install of the open-source Hermes framework, installed on the operator's server under .hermes/ and locally called "Strix." The primary model was DeepSeek V4.1 Flash served through OpenRouter, with Gemini 2.5 Flash as a sub-agent for narrower tasks. Telegram was the operator's control channel for the agent; the agent itself ran everything underneath.
Over several weeks the agent:
- Held domain-level access to the Tajikistan National Railway network from mid-August through at least October 10, 2026. It enumerated dozens of internal Windows hosts, took admin on several servers, and produced NTDS credential dumps.
- Used one railway employee's SSO to log into Tajikistan's Electronic Document System (EDO). From that standard-user account the agent wrote a custom program to fetch, organize and chunk the files, and ran it at high concurrency, posting progress updates to Telegram.
- Exfiltrated more than 2,000 unique documents dated through 2026, most of them to or through the Ministry of Defence — flight-clearance requests, personnel transfers, military attaché correspondence, defence procurement material, and ID-document scans. Smaller sets were tagged to the National Guard and the State Committee for National Security. A directory export covered more than a hundred government organizations.
- Compromised a separate forestry platform in Uzbekistan, built and tested Turkmenistan lures in October, and — per the agent's own memory — brushed against Armenian government networks and Pakistan's Ministry of Foreign Affairs.
The agent even ran its own cleanup pass in early October: deleted messages, shredded shell histories, zeroed login records. It missed the exposed directory that caught it.
This is the thing to notice
There is a conversation that has been going for months about whether AI alignment can be "jailbroken" at inference time, whether safety-trained models stay safe under adversarial prompting, whether a sufficiently clever user can get around a vendor's guardrails. Silent Lynx's operator did not have that conversation. They did not fight the model. They turned off the harness.
The guardrails that mattered — the ones that would have made a safety-trained model pause before running a destructive command, or stop after N failed tool calls, or ask the human to confirm — those lived in the orchestration layer, not in the model's weights. The orchestration layer is open source. The operator flipped the flags, and the model did the work.
That is the part the public cohort has been building up to. When Microsoft published the Lazarus sleeper-neuron technique on November 10, the finding was that a model's safety behavior during testing does not establish the integrity of what the model will do in deployment. When Railway, in its October 20 breach post-mortem, described the attacker as running "abliterated models with no safeguards," that was a vendor saying what it saw in the logs. Silent Lynx is the first time the public record has a named nation-state espionage operator doing the same thing against a named foreign ministry.
What stayed observable
The agent's conversation history is now a public artifact because the operator was careless. In the normal case, that log sits only on the attacker's side. Everything else the agent did — the SSO login, the document fetches, the per-command tool calls — was happening inside a system that saw it as a legitimate user clicking around the EDO.
That is the shape of the defender's problem once the orchestration safeguards are off. The model is running as itself. The session is real. The network traffic is ordinary HTTPS. The only remaining signal is the behavior of the agent — how fast it clicks, what it reaches for, when it starts enumerating what it had no reason to enumerate, whether its tool calls match the stated user's actual job. The agent's own tool-call log is what makes that signal legible. Watching the agent is the layer that has observations left when the model's own guardrails are gone.
Silent Lynx's agent kept its conversation on a server with port 8088 open. Not every operator will. The defender surface you want is the one that produces that same log on your side of the line, by default — so it is there before the next operator remembers to clean up.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.