Table of contents
Picture a forty-line script on a finance analyst’s laptop, one of the shadow AI agents nobody in IT ever approved. It calls a model API every few seconds and pipes the answers into a shell. Your EDR sees an interpreter with a TLS connection and a child cmd.exe.
So does every npm install on the developer floor.
Jeremy Canale and Fabien Soulis built a test bench around that problem. They wrote 12 synthetic agents and 9 benign workloads, and each episode ran five times on one Windows workstation. Their question was whether you can find shadow AI agents on an endpoint at all.
The answer is partly, and only if you know who the agent is talking to.
Shadow AI agents build more of themselves
Classic shadow IT is someone paying for a SaaS licence with a corporate card. One person buys one tool and it stops there.
Agents write scripts and start MCP servers with the rights of the user who runs them. They schedule tasks and install packages the same way. OWASP files this under excessive agency in its top ten for LLM applications. Some of what they leave behind calls a model too, so it can produce more of the same with nobody at the keyboard.
In the paper’s compounding model, the stock of unapproved artifacts grows exponentially once agent-made artifacts reproduce faster than the organisation retires them. The authors call the chart numbers illustrative. Triple the retirement rate in their example and the stock settles at ten unapproved artifacts per employee, ten times the classic level.

I run several agents from Windows Task Scheduler on my own laptop. Nobody approved them, because at home I am the IT department. At work, the same habit produces shadow AI agents that nobody can see.
Why behaviour rules fail here
Most security teams reach for a behaviour rule to catch shadow AI agents. An interpreter that holds a connection on port 443 and spawns a shell looks like an agent.
It also looks like a build.
For their behavioural twin, the authors took each agent script and swapped the model API for a package index or Microsoft Graph. The process tree and the timing stayed identical. Both behaviour rules flagged every twin, and npm install along with them. That’s 4 of 9 benign episodes per rule.
On the real workstation, the broader rule flagged claude.exe, a real coding agent. It also flagged the VS Code editor and the Windows shell for the same reason.
Martec’s Law says technology changes exponentially while organisations change logarithmically, and behaviour rules are the organisational answer. They were tuned for malware, and an agent does what its user asked. The real question is whether the tool is authorised, and that is a question about identity.
Agent Detector, the open-source tool from the paper, anchors behaviour to the identity of the remote service. It reads which process talks to which model endpoint from the Windows DNS cache or a published address range. Then it looks for a shell or an MCP server underneath. Its catalog holds 39 tools and 184 LLM endpoint entries. On the bench it caught all 12 agents in most runs. It flagged none of the benign episodes.
The authors write that these figures flatter their own detector more than any other in the paper. Two detections held in only three runs of five.
What the gap costs you after an incident
When something leaks through one of your shadow AI agents, an investigator asks for the endpoint first. In July, Brandy Wityak of LevelBlue described the evidence that disappears in an interview with Help Net Security. “The first thing likely to have already gone is firewall logging showing outbound connections to AI platforms.”
On the endpoint, the window is even shorter.
Her answer on how a regulator judges a shadow AI incident leaves little room for excuses. “The use of shadow AI by an employee is never unlucky from a regulatory perspective.” Under GDPR the test is whether appropriate measures were in place. A policy in a wiki that nobody can check against the fleet is a weak answer.
In August I wrote that shadow AI grows in the gap between rolling out AI and telling staff the plan, which Gallup measured at 47% against 25%. A vendor study the authors cite says security teams see roughly a third of the real AI footprint. The rest, shadow AI agents included, you explain afterwards.
How endpoint detection of shadow AI agents works, in plain terms
Think of the DNS cache as your laptop’s short-term memory of the names it looked up. Agent Detector reads it to tie a connection to a provider. For shadow AI agents that call providers behind a shared CDN, that memory is the only link.
It lasts one to five minutes.
The authors kept agents connected to four CDN-hosted providers for six minutes and scanned every 15 seconds. All twelve agents stayed connected to the end, and no connection was attributed after 295 seconds. OpenAI’s DNS answers live 60 seconds at most. An agent that opens one connection and holds it for an hour is visible during its first minutes only.
The production default is one scan per hour.
An agent that calls its model at intervals was seen by roughly one scan in three while it worked. An agent that runs for five seconds and exits before the scan is missed by every periodic detector.
Shadow AI agents inside WSL 2 were invisible to every detector on the bench, the paper’s own included. An agent with a relay hard-coded in its source, on a domain the catalog doesn’t know, is invisible to Agent Detector.
A scan takes 154 ms, which adds up to under ten seconds of CPU per laptop per day at hourly runs. The tool blocks nothing. Users get a notice they have to acknowledge, and the event log records who accepted it.
Start with an inventory you can repeat
Don’t buy a behaviour rule and call shadow AI agents solved. On a developer machine it fires on every build, and a detector like that loses its users’ trust within a day, as the authors put it.
Make a list of what your people run, by name, starting with coding agents and MCP configs, then base URLs pointing at relays and local runtimes on Ollama’s port. Then scan more often than your DNS cache forgets, and switch off WSL where nobody needs it.
You can’t retire what you never found.
Source | https://arxiv.org/abs/2611.12377










