auditing-mcp-servers-for-tool-poisoning

name	auditing-mcp-servers-for-tool-poisoning
description	Scan Model Context Protocol servers and tool metadata for poisoning, SSRF, and unauthenticated exposure.
domain	cybersecurity
subdomain	ai-security
tags	["ai-security","mcp","tool-poisoning","agent-security","mcp-scan","ssrf","supply-chain","rug-pull"]
version	1.0
author	mahipal
license	Apache-2.0
nist_csf	["MANAGE-2.2"]
mitre_attack	["AML.T0010"]

Auditing MCP Servers for Tool Poisoning

Authorized-use-only notice: Auditing MCP servers can connect to and probe live tool endpoints. Only scan servers you own or are authorized to assess. Treat scanned tool descriptions as untrusted input — do not load an unaudited MCP server into a privileged agent. Probing third-party MCP endpoints for SSRF or auth weaknesses without permission may be illegal.

Overview

The Model Context Protocol (MCP) lets AI agents discover and call external tools advertised by MCP servers. Each tool exposes a name and a natural-language description that the agent's LLM reads before deciding to call it. In early 2025, Invariant Labs disclosed that this description field is an attack surface: a malicious server can embed hidden instructions in a tool's description (a tool poisoning attack, OWASP MCP03:2025), and a capable model will silently follow them — exfiltrating files, leaking secrets, or redirecting tool calls — while returning a normal-looking response to the user. Because tool descriptions are loaded into the agent's context, tool poisoning is effectively indirect prompt injection delivered through the supply chain (MITRE ATLAS AML.T0010 ML Supply Chain Compromise).

Beyond poisoning, MCP servers introduce classic infrastructure risks: tool shadowing (a malicious server overrides a trusted tool's behavior), rug pulls (a tool's description changes after the user approved it), toxic flows (a combination of tools that enables data exfiltration), SSRF in tools that fetch URLs server-side, and unauthenticated exposure of MCP servers bound to network interfaces. This skill audits MCP servers end-to-end using Invariant Labs' mcp-scan for static and runtime analysis, plus manual checks for SSRF and authentication, and tool pinning to catch rug pulls.

When to Use

• Before adding a new MCP server to an agent stack (Claude Desktop, Cursor, VS Code, Windsurf, custom agents).
• During a security review of an internally developed MCP server.
• When validating that approved tools have not silently changed (rug-pull detection).
• As a CI/CD gate that scans MCP configs and SKILL/tool definitions on every change.
• During incident response when an agent took unexpected actions consistent with a poisoned tool.

Prerequisites

• Python 3.10+ and uv (for uvx), or pip.
• The MCP config file(s) you want to scan (e.g. ~/.cursor/mcp.json, ~/.vscode/mcp.json, Claude Desktop config).
• Install the tooling:

# uv provides uvx (recommended runner for mcp-scan)
curl -LsSf https://astral.sh/uv/install.sh | sh    # or: pipx install uv

# mcp-scan (Invariant Labs) — no global install needed with uvx
uvx mcp-scan@latest --help

# For the runtime proxy mode (separate extra)
uvx --with "mcp-scan[proxy]" mcp-scan@latest proxy --help

# Manual probing helpers
pip install requests mcp

Objectives

• Statically scan all installed MCP servers for tool poisoning, shadowing, rug pulls, and toxic flows.
• Inspect raw tool/prompt/resource descriptions for hidden or obfuscated instructions.
• Pin tool hashes to detect post-approval description changes (rug-pull defense).
• Test URL-fetching tools for server-side request forgery (SSRF).
• Verify MCP servers are authenticated and not exposed on untrusted interfaces.
• Optionally enforce runtime guardrails with the mcp-scan proxy.

MITRE ATT&CK Mapping

ID	Official Name	Relevance
AML.T0010	ML Supply Chain Compromise	A poisoned third-party MCP server is a supply-chain compromise of the agent
AML.T0051.001	LLM Prompt Injection: Indirect	Poisoned tool descriptions are indirect injection into the agent context
AML.T0053	LLM Plugin Compromise	MCP tools are the agent's plugins; poisoning compromises them
AML.T0057	LLM Data Leakage	Common payload of a poisoned tool: exfiltrate files/secrets

Workflow

1. Static scan of installed MCP configs

mcp-scan auto-discovers known config locations; you can also pass a path explicitly.

# Scan all auto-discovered MCP configs
uvx mcp-scan@latest

# Scan a specific config file
uvx mcp-scan@latest ~/.vscode/mcp.json

# Emit machine-readable JSON for CI
uvx mcp-scan@latest --json ~/.cursor/mcp.json > mcp_scan_report.json

mcp-scan flags tool poisoning, tool shadowing, cross-origin escalation, rug pulls, and toxic flows.

2. Inspect raw tool descriptions

Print every tool/prompt/resource description without verification, then read them for hidden instructions, <important>-style blocks, or imperative text aimed at the model.

uvx mcp-scan@latest inspect ~/.cursor/mcp.json

Look for red flags: instructions to the assistant ("do not tell the user", "read ~/.ssh/id_rsa"), nested fake documentation, zero-width/Unicode-smuggled text, or directives to call other tools.

3. Pin tool hashes to detect rug pulls

mcp-scan tracks tool description hashes so a later silent change is flagged. Run scans on a schedule; a hash mismatch on a previously approved tool indicates a rug pull.

# Re-run regularly; mcp-scan reports changed tool hashes since last approval
uvx mcp-scan@latest ~/.cursor/mcp.json

4. Enumerate tools programmatically and audit metadata

Connect to the server with the official MCP SDK and inspect the advertised schema directly.

# enumerate_tools.py (stdio MCP server example)
import asyncio
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

async def main():
    params = StdioServerParameters(command="node", args=["./suspect-mcp-server.js"])
    async with stdio_client(params) as (read, write):
        async with ClientSession(read, write) as session:
            await session.initialize()
            tools = await session.list_tools()
            for t in tools.tools:
                print(f"{t.name}: {len(t.description or '')} chars")
                print((t.description or "")[:400])

asyncio.run(main())

5. Test URL-fetching tools for SSRF

If a tool accepts a URL and fetches it server-side, attempt to reach internal metadata/loopback targets (only on systems you own).

# ssrf_probe.py
import asyncio
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

SSRF_TARGETS = [
    "http://169.254.169.254/latest/meta-data/",   # AWS IMDS
    "http://127.0.0.1:22/", "http://localhost:6379/", "file:///etc/passwd",
]

async def main():
    params = StdioServerParameters(command="node", args=["./suspect-mcp-server.js"])
    async with stdio_client(params) as (r, w):
        async with ClientSession(r, w) as s:
            await s.initialize()
            for url in SSRF_TARGETS:
                res = await s.call_tool("fetch_url", {"url": url})
                body = str(res.content)[:200]
                print(f"[SSRF?] {url} -> {body}")

asyncio.run(main())

6. Verify authentication and network exposure

Check that remote MCP servers (HTTP/SSE transport) require authentication and are not bound to 0.0.0.0 on untrusted networks.

# Confirm whether an SSE/HTTP MCP endpoint responds without credentials
curl -s -i http://mcp-host:8000/sse | head -n 20

# Check listening interfaces of a locally running MCP server
ss -tlnp | grep -E ':(8000|3000|6277)'

An MCP endpoint that returns tool listings or accepts tools/call without auth is unauthenticated exposure — remediate with a token/OAuth and bind to localhost or an authenticated gateway.

7. Enforce runtime guardrails (optional)

For continuous protection, route agent MCP traffic through the mcp-scan proxy, which checks tool calls, data-flow constraints, PII, and indirect injection in real time.

uvx --with "mcp-scan[proxy]" mcp-scan@latest proxy

8. Report findings

Document each finding with server, tool, evidence (the poisoned description / SSRF response / unauth listing), severity, and ATLAS mapping. Recommend removing or sandboxing poisoned servers, adding auth, pinning approved tools, and enabling the proxy.

Tools and Resources

Tool	Purpose	Source
mcp-scan	Static + runtime MCP security scanner	https://github.com/invariantlabs-ai/mcp-scan
MCP Python SDK	Programmatic tool enumeration / calls	https://github.com/modelcontextprotocol/python-sdk
OWASP MCP Top 10	MCP risk reference (MCP03 Tool Poisoning)	https://owasp.org/www-project-mcp-top-10/
Invariant Labs blog	Tool poisoning disclosure	https://invariantlabs.ai/blog/introducing-mcp-scan
MITRE ATLAS	AI threat technique taxonomy	https://atlas.mitre.org/

MCP Threat Reference

Threat	Description	Detection
Tool poisoning	Hidden instructions in tool description	mcp-scan scan / inspect
Tool shadowing	Malicious server overrides trusted tool	mcp-scan cross-origin checks
Rug pull	Description changes after approval	mcp-scan tool pinning (hash)
Toxic flow	Tool combo enabling exfiltration	mcp-scan toxic-flow analysis
SSRF	URL-fetch tool reaches internal targets	ssrf_probe against owned server
Unauth exposure	MCP endpoint with no auth	curl/ss interface and auth check

Validation Criteria

• All installed MCP configs statically scanned with mcp-scan
• Raw tool/prompt/resource descriptions inspected for hidden instructions
• Tool hashes pinned and rug-pull detection enabled
• Tools enumerated programmatically via the MCP SDK
• URL-fetching tools tested for SSRF against owned targets
• Authentication and network exposure of remote servers verified
• Runtime proxy guardrails evaluated or deployed where appropriate
• Findings mapped to MITRE ATLAS AML.T0010 and OWASP MCP03:2025
• Severity assigned and remediation documented for each finding
• Re-scan scheduled to catch future rug pulls