An Agent Left Its Instructions on an Open Server
![clean data visualization, flat 2D chart, muted academic palette, no 3D effects, evidence-based presentation, professional infographic, minimal decoration, clear axis labels, scholarly aesthetic, An open steel vault door hanging ajar on a reinforced concrete wall, spilling hundreds of stamped paper scrolls across a monochrome grid-patterned floor, overhead fluorescent lighting casting sharp perpendicular shadows, atmosphere of sterile abandonment with faint digital timestamp printouts scattered among the papers, one scroll partially shredded by an automated feed slot labeled "Hermes". [fal-ai/z-image/turbo] clean data visualization, flat 2D chart, muted academic palette, no 3D effects, evidence-based presentation, professional infographic, minimal decoration, clear axis labels, scholarly aesthetic, An open steel vault door hanging ajar on a reinforced concrete wall, spilling hundreds of stamped paper scrolls across a monochrome grid-patterned floor, overhead fluorescent lighting casting sharp perpendicular shadows, atmosphere of sterile abandonment with faint digital timestamp printouts scattered among the papers, one scroll partially shredded by an automated feed slot labeled "Hermes". [fal-ai/z-image/turbo]](https://cdn.digitalrain.dev/thelongview/viral-images/85e7b0b8-1bb9-4876-a511-80faffcc3fad_viral_4_square.jpg)
An AI agent executed a targeted intrusion using Hermes, from reconnaissance to exfiltration—yet the operation was only visible because of a misconfigured server. This is a capability signal, not an adoption signal. The distinction matters.
The mistake was a server left open to the internet. Through it, CloudSek, a cybersecurity company based in Singapore, saw what a group it calls Silent Lynx had built around a break-in to Tajikistan's electronic document system. The company says the hackers took more than 2,000 national security-related documents, and its account rests on what it recovered from that server.
The timeline in the recovered data starts in mid-August, when the hackers used Hermes for reconnaissance. A couple of weeks later they were inside the system, and the agent had written a custom program to download the files, organise them and break them into chunks so they could be taken more effectively. “Instead of typing hacking commands by hand, the attacker told Hermes what to do,” CloudSek told Reuters, “and Hermes planned the attacks, wrote the rules, ran them, kept notes, and wrote up reports.”
The instructions found on the server are the most revealing part. They tell the agent how to see itself and how to communicate with the people running it, beginning with the line “You are an elite offensive security specialist and red teamer.” A red teamer simulates covert attacks to test an organisation's systems, so the operators gave the agent a role that resembles authorised testing. The role is a description of intent, not evidence of who authorised the work.
What has changed is narrower than the word autonomy suggests. The operator set an objective, and the agent handled planning, scripting and execution, a faster loop of retrieval and synthesis rather than independent research. The case shows that loop working end to end once. The public evidence cannot yet say how often it works without a careless server to expose it.
Attribution rests on operational habits rather than confirmed identity. CloudSek found that the operators spoke Russian, worked Kazakh business hours and often accessed the server through a residential Kazakh internet connection. Other researchers have tied Silent Lynx to Kazakhstan, but efforts to reach the group were unsuccessful, and Kazakh and Tajik officials had not responded to requests for comment.
Reuters had earlier reported cybercriminals assigning break-ins to autonomous agents. CloudSek says more advanced, espionage-focused hackers are now doing the same. That is one firm's reading of one investigation, and the scale of the practice has not been measured. What made this operation visible was an accident of configuration, not a detection of the agent. The recovered instructions show a tool that was asked to play a professional role across a campaign that began with reconnaissance in mid-August.
† Analysis: the writer's conclusion from the reported facts, not a fact stated in the source.
- The agent's role as a red teamer is a description of intent rather than evidence of authorized testing. The operator set an objective while the agent managed the execution loop, representing a faster retrieval and synthesis process rather than true independent research.
Published November 12, 2026