Tag: shadow ai

  • Shadow AI agents slip out of view in five minutes

    Shadow AI agents slip out of view in five minutes

    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.

    shadow AI agents

    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


    Want More? Subscribe to The Dossier

    Every week in your inbox:

    📡 THE INTELLIGENCE FEED – 3-5 curated links: [Research] [Policy] [Tools] [Incidents]
    💡 ONE ADVICE – One actionable AI/cybersecurity tip you can use today

  • Shadow AI grows in the gap Gallup just measured

    Shadow AI grows in the gap Gallup just measured

    Table of contents

    Ask a US employee whether their own employer has rolled out AI. A decent share of them cannot answer, and shadow AI grows in exactly that kind of confusion.

    Gallup changed the question because of it. The methodology note says, “Starting in Q3 2025, Gallup added a ‘don’t know’ option to this question to capture uncertainty about AI adoption.” Results from Q3 2025 onward are no longer directly comparable with earlier measurements.

    A polling company broke its own time series because too many people had no idea what was happening inside the building they work in.

    The second number nobody quotes

    As of May 2026, 47% of US employees say their organization has implemented AI. Only 25% say the organization communicated a clear plan.

    Every headline takes the first number, but the second one is the story.

    Between “we have this thing” and “somebody told me how to use it” sits a gap, and in that gap are contracts, patient records, draft tenders, source code, whatever your people happen to be pasting today.

    Shadow AI is the default state of that gap.

    shadow AI

    Adoption is the wrong argument

    There is a long running fight about how many people use AI. Gabriel Weinberg of DuckDuckGo summed up the skeptical side in June 2026 as “one third actively using AI, one third occasionally using AI, and one third never using AI”. He cites Microsoft telemetry putting it at “more than 30 percent of the US working-age population is using AI, an increase of 3 percentage points from the end of 2025”.

    Gallup’s workplace numbers run higher. 15% of US employees use AI daily. Weekly or more is 30%, and 52% touch it at least a few times a year.

    Pick whichever camp you like, it changes nothing for my job.

    Move the user base up or down, the 25% who got a clear plan stays where it is. That fight pulls attention away from the only question that matters, which is who wrote the rulebook and who read it.

    Shadow AI is a confidentiality problem with no attacker

    Strip the vocabulary and that is all this is.

    No phishing mail, no exploit, no command and control, nothing that trips an alert. An employee opens a browser tab and pastes a client document into a chatbot to get a summary. The data leaves the organization. Most of what you bought assumes somebody is trying to break in. Shadow AI walks out the front door during working hours, moved by people who just want to finish faster.

    OWASP keeps an entry for sensitive information disclosure in its Top 10 for Large Language Model Applications, and almost all of that guidance assumes an application you built. The tab your sales team opened this morning is nobody’s application.

    Scott Brinker named the shape of this back in 2013 and called it Martec’s law. Technology changes exponentially, organizations change logarithmically. Your staff adopted AI in an afternoon, your document set moves at the speed of a committee.

    The Gallup manager numbers show the same thing from the other side. 36% strongly agree their manager supports the team using AI. Where that support exists, employees are 1.7x more likely to use AI weekly or more and 8.7x more likely to report a transformational change in how they work. Encouragement travels by conversation and rules travel by document. Shadow AI takes the faster route.

    What shadow AI looks like on a Tuesday

    Nobody sits down and decides to run shadow AI. It shows up as small, reasonable moves.

    A sales rep pastes a signed contract into a chatbot to pull the renewal dates out of it. An HR assistant drops a salary spreadsheet into a chatbot to reformat the columns. A developer sends a stack trace holding a production connection string to a free tier account. A clinic receptionist rewrites a referral letter with the patient name still in it.

    None of those people are careless. All of them were told AI makes them faster, and none of them were told where the line sits.

    Deleting the client name before pasting does not turn the text into anonymous data either, and I went through the research on that in ChatGPT privacy leak.

    Every one of those actions is invisible to the security team, because nothing was breached and nothing alerted.

    Low numbers are not safe numbers

    Daily use runs at 42% in technology, 27% in finance, 22% in professional services, and 9 to 15% everywhere else.

    Read the low end carefully, because a law firm or a clinic sitting in that bottom band is not in a better position. It is a place where a smaller group does the same thing with far more sensitive material, and with less chance that anyone in IT has ever looked at it. Low usage hides shadow AI.

    The 65% of employees in AI-implementing organizations who report a positive effect on productivity are not lying either. It works, and that is exactly why nobody is going to stop when you ask them to. A ban moves shadow AI further out of sight.

    One page before you buy anything

    Do not start with a tool. Start with one page that answers four things.

    1. Which categories of data never go into an external model.
    2. Which tools are approved, listed by name.
    3. Who an employee asks when the answer is not obvious.
    4. What happens when something has already gone in, and who hears about it first.

    That page will not cover a regulated environment or replace a contract with the vendor. It covers what is leaking today.

    I keep the editable template for it behind my shadow AI risk calculator, so you do not have to start from a blank file.

    Then do the boring part and send it to everyone, with one person named as the owner. A rule nobody can find works the same as no rule at all, and that is where shadow AI restarts. It is not fancy work and it does not need a consultant.

    One page beats a procurement cycle. If you cannot write it, you do not have an AI program, you have 47% and hope.

    So remember, if you write only one line on that page, write this one. Never put anything into an LLM that you would not email to a stranger.

    Source: https://www.gallup.com/699797/indicator-artificial-intelligence.aspx


    Want More? Subscribe to The Dossier

    Every week in your inbox:

    📡 THE INTELLIGENCE FEED – 3-5 curated links: [Research] [Policy] [Tools] [Incidents]
    💡 ONE ADVICE – One actionable AI/cybersecurity tip you can use today