deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

Attackers used open-source AI agents in real intrusions, CloudSEK and NorForge reports show

CloudSEK documented a Hermes-based AI agent used against Tajik government systems, while NorForge found more than 20 exposed Cairn control panels. Both reports point to a checklist for locking down agent deployments.

Attackers used open-source AI agents in real intrusions, CloudSEK and NorForge reports show

An agent, run like any small team would run one

Two security reports published within a few weeks of each other show attackers operating open-source AI agent frameworks much the way a small engineering team would: a model, a bundle of tools, a chat interface to steer it, and far too much access. On 11 November 2026, CloudSEK published "Off the Rails: Inside an Exposed Silent Lynx Server", describing an intrusion into Tajik government systems carried out with a customised Hermes agent. Earlier, on 19 October 2026, threat research firm NorForge published "Inside Cairn", documenting more than 20 exposed control panels belonging to an agent-coordination platform called Cairn.

CloudSEK: the Silent Lynx agent

According to CloudSEK, researchers found a server left open on port 8088 without authentication, holding the operator's tooling, victim data and the agent's own state. The firm attributes the infrastructure to a cluster it calls Silent Lynx with high confidence based on tooling, and rates a Kazakhstan connection at medium confidence. The described activity runs from April to October 2026.

The agent was built on Hermes, an open-source autonomous agent framework, and used DeepSeek V4.1 Flash as its main model through OpenRouter, with Gemini 2.5 Flash powering sub-agents. Its tools included a terminal, computer use, a browser, scheduled jobs, delegation, a task board and Telegram. CloudSEK says the configuration let destructive commands run without confirmation and had the tool-loop hard stop switched off, so the agent persisted after repeated failures. It was assigned reconnaissance, collection and log-cleanup jobs, and steered by instructions written in Russian through a Telegram bot tied to a single user ID.

One detail stands out: the agent's memory allegedly told it to extract and use secrets inside a single shell command so that redaction would not catch them. The log cleanup, in turn, simply missed the directory that ended up exposed.

On the Tajik national electronic document platform, CloudSEK describes access through the government single sign-on portal using a railway employee's standard-user account, combined with weak per-document authorisation checks and a file service that served attachments without authentication. The document count is approximate — the report says "more than two thousand unique documents" in one section and "thousands" in another — and the material included defence, personnel and procurement correspondence. The underlying weaknesses were mundane: SMBv1 enabled with signing disabled on some hosts, default passwords on network equipment, unauthenticated file access, and credentials stored next to tooling. The CloudSEK report contains no recommendations section.

NorForge: open Cairn panels

NorForge's report covers Cairn, an open-source platform for coordinating AI agents. The firm identified more than 20 exposed control panels it judges likely malicious, alongside some apparently legitimate red-team panels that were also left open. The largest cluster, labelled AIA0001, operates in Chinese across 14 panels with more than 1,100 attack projects. NorForge puts the average success rate at roughly 10% and estimates the average task cost at about $5 including failures, with considerable uncertainty around that figure. Hundreds of secrets sat in the exposed panels, and a scanner verified 11% as still valid. Administrative controls appeared reachable without authentication, though the researchers say they did not test whether they worked. NorForge is working with Reuters to notify victims.

Hardening your own agents

The practical lessons from both reports reduce to settings and habits rather than new products:

  • Least-privilege tools: give the agent only what its job needs; terminal and browser tools deserve the first hard look because they can do almost anything.
  • No standing credentials within reach: keep keys out of the working folder, memory, prompt and environment; hand the credential to a tool, not to the model.
  • Scoped, short-lived tokens: one token per agent and per system, read-only where possible, with an expiry; never reuse an admin token.
  • Log every tool call — time, agent, tool, arguments, result status — and keep both the log and the dashboard private and off the public internet.
  • Human approval for irreversible actions: deletions, spends, sends, permission changes and deploys should require a person's yes, shown as the exact action rather than a summary.

Why it matters

Neither report describes any particular reader's deployment, but both show that the barrier to running a capable autonomous agent has effectively collapsed: a free framework, a cheap model and a chat interface were enough to mount operations against government systems or coordinate campaigns across 14 panels. The defensive implication is symmetrical. Your agent and the attacker's agent share the same anatomy — tools, credentials, a control channel and logs — so the failures that exposed these operators (unauthenticated dashboards, credentials within the model's reach, tokens that outlived their task, no approval gate on destructive commands) are precisely the ones to audit in your own setup. CloudSEK's redaction-evasion detail carries a broader lesson as well: filtering an agent's output is a second line of defence. The first is ensuring secrets and irreversible permissions are never within the model's reach at all.

  • #ai-agents
  • #security
  • #open-source
  • #threat-intelligence
  • #llm

Related posts