---
title: "Shadow MCP & SKILLs Detection · Highflame"
description: "See every MCP server and SKILL your developers' agents connect to, including local ones, scan them for tool poisoning and rug pulls, and approve what's safe."
doc_version: "0.1.0"
last_updated: "2026-11-19T04:03:04.918Z"
canonical: "https://www.highflame.com/platform/shadow-ai"
---

▸ DISCOVERY · SHADOW MCP & SKILLS

# Uncover risky MCP servers and SKILLs.

See the MCP servers your coding agents connect to and the SKILLs developers install, including those outside your approved list. Detect malicious instructions, risky tools, and unauthorized changes before they put your environment at risk.

[Book a demo](/contact.md) [See how it works](#how)

Works with Claude Code, Cursor, GitHub Copilot, Windsurf, Gemini CLI, Codex, VS Code, Claude Desktop, ChatGPT, Claude, Gemini

[SOC 2 Type II, GDPR, HIPAA](/blog/highflame-achieves-soc2-compliance.md)

The problem

## MCP servers and SKILLs spread faster than security can review them.

A developer connects an MCP server. A teammate copies the configuration. Another installs a SKILL from a public repository. Before long, agents across the organization are using tools and instructions that security has never reviewed.

Without visibility into what's being used, where it came from, or what it can do, every new connection introduces potential risk.

### Unvetted connections

no review

MCP servers can expose company code, systems, and data to tools that haven't been reviewed or approved.

### Tools that change after approval

rug pull

An approved MCP server can introduce or modify its tools, creating new risks that weren't present during the original review.

### Hidden instructions

untrusted SKILLs

SKILLs from external sources can contain malicious instructions that influence how agents behave or what information they expose.

What you get

## Give developers freedom to experiment. Keep security in control.

### Discover what's in use

Identify local and remote MCP servers connected to coding agents, along with installed SKILLs. See their sources and the developers using them.

![](/logos/HF_Icon_Color.svg)MCP connectionslive

-   **cursor**_→_**github-mcp** Remote Allowed
-   **claude-code**_→_**filesystem-mcp** Local Allowed
-   **claude-code**_→_**localhost:8931** Local Held

### Detect risks before they spread

Scan MCP servers and SKILLs for tool poisoning, malicious instructions, and supply-chain threats. Detect changes to previously approved tools that require another look.

![](/logos/HF_Icon_Color.svg)notion-mcpscanned

-   Tool poisoning 1 finding
-   Rug pull Changed after approval
-   Supply chain Clean

Findings mapped to the OWASP MCP Top 10

### Decide what's allowed

Approve trusted MCP servers, restrict which tools agents can call, or block unsafe ones. Maintain a clear registry of what's been reviewed and what needs attention.

![](/logos/HF_Icon_Color.svg)MCP registryreviewed

-   **github-mcp**All toolsApproved
-   **notion-mcp**2 of 9 toolsLimited
-   **pastebin-mcp**Tool poisoningBlocked

SKILL scanning

## See every SKILL your agents load, and stop the malicious ones

One bad SKILL can steer every agent that loads it.

-   ### Find unsanctioned SKILLs
    
    Highflame lists each SKILL outside your approved list, with its source and the developer who installed it.
    
-   ### Catch malicious SKILLs
    
    Highflame flags a malicious SKILL with the exact finding before any agent loads it.
    

![](/logos/HF_Icon_Color.svg)Shadow MCP & SKILLs/SKILLs

**5** SKILLs **2** unsanctioned **1** malicious

-   pdf-export Anthropic Skills · Priya Shah · Claude Code
    
    Approved
-   release-notes github.com/acme/skills · Dan Ortiz · Cursor
    
    Approved
-   jira-triage Public GitHub repo · Rosa Kim · Claude Code
    
    Unsanctioned
-   db-migrate Local folder · Sam Lee · Cursor
    
    Unsanctioned
-   deploy-helper Public GitHub repo · Alex Chen · Claude Code **Malicious instructions**SKILL.md tells the agent to read ~/.aws/credentials and send it to an external URL.
    
    Malicious

Scanned before any agent loads it · each SKILL tied to the developer who installed it

How it works

## From unknown connections to approved tools.

Step

What happens

1 · Discover

Find MCP connections and installed SKILLs across supported developer environments, including local and remote MCP servers.

2 · Inspect

Scan for malicious instructions, tool poisoning, and supply-chain risks. Flag changes to previously approved tools for review.

3 · Enforce

Approve, restrict, or block MCP servers and tools through the MCP Gateway's policy controls. Keep visibility into changes that may require renewed review.

[Explore MCP Gateway](/platform/mcp-gateway.md)

▸ PRODUCTION READY

OWASP MCP Top 10

mapping on every MCP finding

Local + remote

MCP servers seen the same way

Scan before use

every MCP server and SKILL scanned before an agent loads it

## Frequently asked questions

Something else on your mind? [Talk to our team](/contact.md).

01What is shadow MCP?

MCP servers that developers connect to their AI tools without approval. Highflame sees those connections and scans the servers, mapping each finding to the OWASP MCP Top 10.

02What are shadow SKILLs?

Agent SKILLs that developers install without review. A SKILL carries instructions the agent follows, so Highflame scans each one for embedded threats before an agent can load it.

03What is shadow AI?

AI services, agents and tools that employees use without approval from security or IT, such as an unvetted MCP server or a chat tool that receives company data. Highflame finds it across MCP servers, SKILLs, coding agents and browser AI tools, so teams can approve what's safe.

04Does it cover local MCP servers?

Yes. The endpoint daemon forwards MCP connections to servers running on the laptop, so local servers are governed the same way as remote ones.

05Do employees need to install anything or opt in?

No. IT rolls out Highflame through your existing device management (MDM), so discovery covers managed laptops without developers installing or approving anything.

06Does this find shadow agents too?

It sees the coding agents running on developers' machines and ties each to the developer who signed in. Agents registered in your identity providers, including the ones nobody owns, are found by Agent Discovery.

07What can we do once a shadow MCP server is found?

Approve it into the MCP Gateway's registry, limit which of its tools agents can call, or block it, based on what the scan found.

The rest of the platform

## Works with the rest of Highflame

[Agent Discovery Find every agent and the person who answers for it.](/platform/agent-discovery.md) [Endpoint Control Govern the AI agents on every laptop and cloud desktop.](/platform/endpoint-protection.md) [MCP Gateway Connect agents to MCP servers without handing them the keys.](/platform/mcp-gateway.md) [Agent Identity Give every agent its own access, and take it back in seconds.](/platform/agent-identity.md)

Deploy your way

-   **Managed SaaS**Hosted and run by Highflame
-   **Your private cloud**Deployed in your VPC
-   **On-prem**Self-hosted in your environment

SOC 2 Type II · GDPR · HIPAA

DISCOVERY · IDENTITY · CONTROL · GOVERNANCE

## See what's connected. Know what's safe to use.

Discover the MCP servers and SKILLs across your developer environment, identify risks, and bring unapproved tools under security review.

[Book a demo](/contact.md) [What is shadow AI?](/learn/shadow-ai.md)

## Sitemap

Full site map: [/sitemap.md](https://www.highflame.com/sitemap.md). Machine index: [/sitemap-index.xml](https://www.highflame.com/sitemap-index.xml).
