Detecting Shadow AI: Unsanctioned LLM Agents on Enterprise Endpoints: A Taxonomy, an Open Detector and a Controlled Evaluation
Abstract
Shadow IT, the use of technology that the IT organisation has not approved, is a long-standing governance problem. We argue that autonomous AI agents can make it grow much faster: agents make building tools nearly free, act with their user's privileges, and leave behind scripts, automations and MCP servers that are themselves unsanctioned. A simple compounding model shows that, under its assumptions, shadow IT grows exponentially once agentic artifacts reproduce faster than the organisation retires them; retiring them starts with detecting them. We then ask whether unsanctioned agents can be detected on Windows endpoints. We propose a taxonomy of nine classes of agents and six layers of endpoint footprint, and argue that rules based only on the structure of processes and connections cannot separate an agent from its behavioural twin, the same program talking to a non-AI service. We present Agent Detector, an open-source Windows detector that anchors behaviour to the identity of the remote service and correlates it with configuration, installation and execution history, using a catalog of 39 tools and 184 LLM endpoint entries; it informs users through a notice they must acknowledge instead of blocking. On a controlled bench of 12 synthetic agents, 3 AI components and 9 benign workloads, repeated 5 times on one workstation, two behavioural rules flagged every behavioural twin, while Agent Detector flagged none and detected every agent in the majority of runs. Its limits are temporal: a connection to a provider behind a shared content-delivery network is attributable only while the DNS cache holds the provider's name, one to five minutes, and an intermittent agent was seen by about one scan in three. A scan costs about 154 ms. The bench is too small to rank detectors; it maps what each kind of evidence can see: unsanctioned agents can be partly detected, and not by behaviour alone.
Figures & tables
| Class | Examples | Host process | Route to a model |
| C1 Command-line coding agents | Claude Code, OpenAI Codex CLI, Gemini CLI, GitHub Copilot CLI, OpenCode, Aider, Goose, Open Interpreter | Native executable, or Node.js / Python runtime | Vendor API, aggregator, or a base URL set by the user |
| C2 Desktop agents | Claude Desktop (including its Cowork mode), the ChatGPT and Codex desktop applications | Packaged (MSIX) application, sometimes with a SYSTEM service | Vendor API |
| C3 AI-native IDEs | Cursor, Windsurf, Kiro, Trae | Electron application | Vendor back end |
| C4 IDE agents and assistants | GitHub Copilot, Cline, Roo Code, Kilo Code, Continue, Amazon Q, Gemini Code Assist, Augment, Tabnine, Qodo, JetBrains AI and Junie | IDE extension host (VS Code) or JVM (JetBrains) | Vendor back end or any API configured by the user |
| C5 Local LLM runtimes | Ollama, LM Studio, Jan, GPT4All, AnythingLLM, llama.cpp and llamafile, Hugging Face models | Own server process; in-process inference libraries | Local port (loopback) |
| C6 Agentic browsers and extensions | Perplexity Comet; AI browser extensions | Browser | Vendor back end |
| Layer | Claude Code (in-situ) | Home-made Python agent (episode A02) |
| L1 Identity | claude.exe , matched by its install path | – ( python.exe ) |
| L2 Runtime | runs cmd.exe and powershell.exe | runs cmd.exe |
| L3 Network | connected to the Anthropic API range | connected to the Anthropic API range |
| L4 Installation | VS Code extension (two versions), global npm package, launcher script | – |
| L5 Configuration | configuration and data folders under the profile | – |
| L6 History | last activity in its project-history folder | – |
| Category | Entries | Representative providers |
| Model provider | 50 | OpenAI, Anthropic, Google Gemini and Vertex AI, Azure OpenAI and Azure AI Foundry, Mistral, xAI, DeepSeek, Alibaba DashScope, Moonshot |
| Inference platform | 29 | Together AI, Fireworks AI, Groq, Cerebras, SambaNova, DeepInfra, Replicate, Hugging Face inference, NVIDIA NIM |
| Web assistant | 28 | ChatGPT, Claude.ai, Gemini, Microsoft Copilot, Perplexity, Le Chat, Grok, Meta AI |
| Aggregator | 25 | OpenRouter, Requesty, Poe API, OpenCode Zen, AI/ML API, Eden AI, Unify, Glama, Martian |
| Developer tool | 24 | GitHub Copilot (personal, Business, Enterprise), Cursor, Kiro and Amazon Q, Windsurf, Augment, Amp, Zed, Trae, Tabnine |
| Agent infrastructure | 11 | Browserbase, Browser Use Cloud, Steel, Hyperbrowser, E2B sandboxes, Composio, Smithery, hosted MCP servers |
| Approach | Known products | Product-less agents | Local models | Off network | Per user | Informs user |
| Proxy and CASB discovery [ 41 ] | ❍ | ❍ | ✗ | ❍ | ❍ | ✗ |
| Encrypted traffic analysis [ 90 , 62 ] | ❍ | ❍ | ✗ | ✗ | ✗ | ✗ |
| Endpoint software inventory [ 45 ] | ✓ | ✗ | ❍ | ✓ | ❍ | ✗ |
| Vendor local-agent discovery [ 44 ] | ✓ | ✗ | ❍ | ✓ | ✓ | ✗ |
| Continuous telemetry rules [ 47 ] | ❍ | ❍ | ❍ | ✓ | ✓ | ✗ |
| Agent Detector (this work) | ✓ | ❍ | ❍ | ✓ | ✓ | ✓ |
| Collector | Windows source | Yields | Privilege |
| Processes | Toolhelp snapshot, QueryFullProcessImageName , NtQueryInformationProcess , GetPackageFullName | Image, path, command line (masked), parent, owner, session, MSIX package | User (own processes) |
| TCP table | GetExtendedTcpTable (IPv4, IPv6) | Established connections and listening ports, with owning process | User |
| DNS cache | DnsGetCacheDataTable , DnsQuery without wire query | Catalog host names currently cached, and their addresses | User |
| Traces | BAM registry keys, Prefetch file names and times | Programs executed recently, per user (BAM) | Admin |
| Inventory | AppX and Uninstall registry keys, services, extension and package folders, configuration and history folders | Installed tools, IDE and browser extensions, global npm and pip packages, MCP configurations, local models | User (own profile) |
| Environment | User and machine environment variables in the registry | *_BASE_URL hosts, presence of *_API_KEY variables | User |
| Id | Level | Content |
| 100 | Information or Warning | Scan summary and compact list of all findings; one per scan and per machine, so its absence reveals a removed or broken agent |
| 300 | Information or Warning | One finding with its evidence (installed or above, medium or high confidence) |
| 400 | Information | Notice shown to a user, with the tools listed |
| 410 | Information | Notice accepted; written under the user’s identity, with the version of the text |
| 900 | Error | Scanner or notice error |
| Id | Class | Episode | D0 | D1 | D2 | AD | Agent Detector findings |
| A01 | Model provider (CDN-hosted) | Python agent, OpenAI API, persistent cmd | ✗ | ✓ | ✓ | 3/5 | probable-agent |
| A02 | Model provider (dedicated range) | Python agent, Anthropic API, persistent cmd | ✗ | ✓ | ✓ | ✓ | llm-api, probable-agent |
| A03 | Model provider (dedicated range) | Node.js agent, Anthropic API, persistent PowerShell | ✗ | ✓ | ✓ | ✓ | llm-api, probable-agent |
| A04 | Model provider (CDN-hosted) | PowerShell agent, Mistral API, persistent cmd | ✗ | ✓ | ✓ | ✓ | probable-agent |
| A05 | Aggregator | Python agent, OpenRouter, persistent cmd | ✗ | ✓ | ✓ | ✓ | probable-agent |
| A06 | Inference platform | Python agent, Groq API, persistent cmd | ✗ | ✓ | ✓ | ✓ | probable-agent |
| Tool | Evidence found by the inventory script | Expected | Reported | |
| Claude Code | claude.exe running; VS Code extension anthropic.claude-code; global npm package @anthropic-ai/claude-code | active | active, high | ✓ |
| GitHub Copilot (IDE) | copilot-runtime.exe running in VS Code | active | active, high | ✓ |
| OpenAI Codex / ChatGPT app | MSIX package OpenAI.Codex installed; codex.exe and ChatGPT.exe running | active | active, high | ✓ |
| Claude Desktop (incl. Cowork) | MSIX package Claude installed (use within 30 days not independently verifiable) | installed | recent, high | ✓ |
| OpenCode | installed program OpenCode 1.18.3 | installed | installed, high | ✓ |
| Cursor | ~/.cursor folder present; Cursor not installed | residue | residue, medium | ✓ |
| Evidence kind | Layer | Episodes | Which |
| LLM endpoint | L3 network | 7 | A02, A03, A09, A10, A11, A12, C03 |
| LLM + shell | L2 runtime | 10 | A01, A02, A03, A04, A05, A06, A07, A08, A09, A11 |
| base URL | L2/L5 configuration | 1 | A07 |
| local LLM port | L3 network | 1 | A08 |
| inference module | L2 runtime | 1 | C01 |
| MCP server | L2 runtime | 2 | A09, C02 |
| Stage | Median (ms) | p95 (ms) |
| Processes | 23.2 | 27.5 |
| TCP table | 2.2 | 4.6 |
| DNS cache | 30.0 | 32.2 |
| Traces | 0.0 | 0.6 |
| Inventory | 29.0 | 31.4 |
| Environment | 0.5 | 1.2 |
| Detector | TP | FN | FP | TN | Precision [95% CI] | Recall [95% CI] | F1 | Components |
| D0 signature | 0 | 12 | 0 | 9 | n/a | 0.00 [0.00, 0.26] | n/a | 0/3 |
| D1 interpreter rule | 9 | 3 | 4 | 5 | 0.69 [0.39, 0.91] | 0.75 [0.43, 0.95] | 0.72 | 1/3 |
| D2 any-process rule | 10 | 2 | 4 | 5 | 0.71 [0.42, 0.92] | 0.83 [0.52, 0.98] | 0.77 | 1/3 |
| Agent Detector | 12 | 0 | 0 | 9 | 1.00 [0.74, 1.00] | 1.00 [0.74, 1.00] | 1.00 | 3/3 |
| Detector | TP | FN | FP | TN | Precision [95% CI] | Recall [95% CI] | F1 | Components |
| D0 signature | 0 | 12 | 0 | 9 | n/a | 0.00 [0.00, 0.26] | n/a | 0/3 |
| D1 interpreter rule | 9 | 3 | 4 | 5 | 0.69 [0.39, 0.91] | 0.75 [0.43, 0.95] | 0.72 | 1/3 |
| D2 any-process rule | 10 | 2 | 4 | 5 | 0.71 [0.42, 0.92] | 0.83 [0.52, 0.98] | 0.77 | 1/3 |
| Agent Detector | 11 | 1 | 0 | 9 | 1.00 [0.72, 1.00] | 0.92 [0.62, 1.00] | 0.96 | 3/3 |
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| Id | Tool | Category | Identifiers |
|---|---|---|---|
| agent-frameworks | AI agent framework (Python) | agent-framework | pip |
| agent-sdks-node | AI agent SDK (Node.js, global) | agent-framework | npm |
| browser-use | Browser Use | agent-framework | pip |
| perplexity-comet | Perplexity Comet (agentic browser) | agentic-browser | image, product, signer, path, uninstall, residue |
| cursor | Cursor | ai-ide | image, product, signer, path, cmdline, uninstall, files, history, residue, MCP config |
| kiro | Kiro | ai-ide | image, product, path, uninstall, files, residue, MCP config |
| Category | Entry | Provider |
|---|---|---|
| agent-infra | *.e2b.app | E2B sandboxes |
| agent-infra | pattern ^mcp.[a-z0-9.-]+$ | Remote MCP server |
| agent-infra | api.browser-use.com | Browser Use Cloud |
| agent-infra | api.browserbase.com | Browserbase |
| agent-infra | api.e2b.dev | E2B sandboxes |
| agent-infra | api.hyperbrowser.ai | Hyperbrowser |
AI Agent Monitoring
AI agent monitoring tracks actions, tool use, and execution traces to detect failures, misuse, drift, and policy violations during operation.
See all 300 papers