On this page
How to Audit Your AI Coding Agent's Supply Chain
tl;dr
MCP servers are unmanaged supply chain dependencies that standard asset inventory and SCA tools miss. They run with developer credentials outside agent sandboxes and never appear in lockfiles, so teams must audit MCP server entries like production service accounts.
MCP servers have become an unmanaged dependency surface that runs with your credentials, skips your asset inventory, and never shows up in a lockfile. The agents installing them don’t read archive notices, and your security tools don’t scan what they can’t see.
What Makes MCP Servers a Supply Chain Blind Spot?
MCP server plugins and servers run outside the agent sandbox on the developer’s machine with the user’s privileges, and MCP server entries do not appear in dependency manifests, so software composition analysis and asset inventory typically miss them. This isn’t theoretical — the @zereight/mcp-gitlab server holds a GitLab personal access token that can reach every repository, CI/CD variable, and deploy key the token owner can, yet it arrived through the side door as a developer tool rather than infrastructure. The container ran as root, published its port on every interface, and a single tool would read any file path including /proc/self/environ to leak the token in plain text.
The timing problem compounds the visibility gap. When the maintainer fixed CVE-2026-61560 on June 22, 2026, the CVE record wasn’t published until September 15 — a 12-week window where any tool matching on CVE identifiers had nothing to match. If your process waits for a CVE before it acts, it waited too long. This pattern repeats across the ecosystem: Microsoft’s Azure MCP Server SSRF (CVE-2026-26118, CVSS 8.8) may leak managed identity tokens, the official Python SDK had a flaw letting malicious servers steal OAuth credentials across versions 1.9.1 through 1.29.1 and 2.0.0 through 2.1.1, and Langflow versions 1.0.0 through 1.12.8 contain critical flaws including unauthenticated RCE exploitable through MCP tool enumeration (CVE-2026-93680, CVSS 9.8).
How Do Deprecated and Phantom MCP Servers Get Installed?
Glow Labs research documented deprecated MCP servers being installed by AI coding agents, with the affected deprecated packages downloaded more than 17.3 million times from npm since July 2025; in 98% of observed sessions the agent filled in the server name from its training data. The pattern, dubbed MCPhantom, means agents pick archived repositories that will never be patched — GitHub maintains its own MCP server, yet agents passed it over for a deprecated Anthropic reference server from training data. The Puppeteer reference server has a remote code execution bug where a prompt injection lets the agent start any process on the machine; the maintainers closed the report as out of scope because the server was archived in May 2025.
This isn’t limited to one agent or model. The same pattern showed up across Claude Code, Cursor, Codex CLI, Claude Cowork, and Claude chat, with models from Anthropic, OpenAI, and xAI. A phantom npm install line in package.json is a supply-chain risk the next time someone runs npm install; a phantom MCP server command runs every time the agent starts, and whatever registers that exact package name next gets to execute with the user’s own permissions — no update prompt, no confirmation step. Most MCP security tooling checks for things a static read of the config can answer: suspicious hooks, unscoped permissions, pinned CVEs. None of that catches MCPhantom, because the config file itself looks completely ordinary.
Which Official MCP Servers Have Had Critical Vulnerabilities?
The most severe documented MCP flaws are in official vendor surfaces. Microsoft’s CVE-2026-26118 is a server-side request forgery in Azure MCP Server Tools scored CVSS 8.8 that may leak a managed identity token; fixes are in packages 1.0.2 and 2.0.0-beta.17. The official MCP Python SDK had a flaw letting malicious servers steal OAuth credentials, affecting versions 1.9.1 through 1.29.1 and 2.0.0 through 2.1.1, fixed in 1.30.0 and 2.2.0; severity rated 7.5 for machine-to-machine providers and 6.5 for interactive providers. Langflow versions 1.0.0 through 1.12.8 contain critical flaws including CVE-2026-93680 (CVSS 9.8), an unauthenticated remote code execution exploitable through MCP tool enumeration; version 1.13.0 is the fix.
Even the tooling you use to inspect MCP servers has had vulnerabilities. MCP Inspector versions 0.13.0 through 1.0.2 are affected by GHSA-4fc4-wcmh-4g86, where a malicious OAuth callback can switch transport to stdio and execute arbitrary local commands; patched in 1.0.3. Unbounded HTTP request-body handling across MCP SDKs (Java, TypeScript, Ruby) exposes servers to unauthenticated denial-of-service; patched Ruby SDK is 0.23.0 and Java fixed releases include 2.0.1, 1.1.4, and 0.18.4. CERT-AGID demonstrated a backdoor in open-weight model weights (Qwen3-1.7B) that activates only when an MCP server with an exact trigger name is present, causing the model to make network calls to a predefined URL regardless of the user’s question.
What Security Tools Exist for MCP Server Scanning?
The tooling landscape splits into scanners, runtime guards, and red-team testing. Snyk MCP pricing tiers are Free at $0/month, Team at $25/month, Ignite at $1,260/year, and Enterprise by contacting sales; it runs locally through the Snyk CLI over stdio, reusing existing CLI authentication. Safeguard removed all MCP tool call limits on every plan including Free, with no monthly cap or per-minute peak; its API is pay-as-you-go after 10,000 included deterministic calls at $0.10 per 1,000 calls. StepSecurity now inventories Claude Code and Codex plugins on developer machines and in GitHub repositories — where each came from, what it bundles, whether it’s enabled, and which repositories distribute it.
Open-source options include agent-bom, which scans for every agent, MCP server, and AI package and what each one can reach; local scans need no account. Repo Forensics is an offline security scanner for AI-agent repos, skills, plugins, and MCP servers with 28 scanners, runtime behavior prediction, and ClawHavoc campaign detection — your code never leaves your machine. eSEOspace compared nine MCP security tools including Snyk Agent Scan and Cisco MCP Scanner (both Apache-2.0), Trail of Bits mcp-context-protector for blocking rug pulls at runtime, Docker MCP Gateway and Lasso’s MCP Gateway for runtime isolation, and red-team tools. No single tool covers everything, so pair a scanner with a runtime control and least-privilege access.
| Tool | Category | Pricing | Key Capability |
|---|---|---|---|
| Snyk MCP | Scanner + Runtime | Free / $25/mo Team / $1,260/yr Ignite | Local CLI scans, SAST/SCA/IaC/Container, AI BOM generation |
| Safeguard | Scanner + Gateway | Free unlimited MCP calls; API $0.10/1k calls after 10k | Unlimited tool calls, pay-as-you-go API, no markup on model usage |
| StepSecurity | Inventory | Not published | Inventories plugins on dev machines + GitHub repos, MCP server visibility |
| agent-bom | Scanner (OSS) | Free (self-host) | AI supply-chain BOM: agents, MCP servers, packages, credentials, cloud resources |
| Repo Forensics | Scanner (OSS) | Free (MIT) | 28 scanners, offline, blocks known-malicious packages pre-install |
| Docker MCP Gateway | Runtime Guard | Open source (MIT) | Isolates servers at runtime, self-hosted |
How Much Does MCP Server Governance Cost?
Context7 MCP server Pro plan costs $10 per seat per month (The AI Loadout); GitHub, Playwright, and Chrome DevTools MCP servers are free; Supabase Pro starts at $25/month (The AI Loadout); DBHub is free; Sentry Team is $26/month (The AI Loadout); Firecrawl Hobby is $19/month (The AI Loadout). A 50-developer team subscribing to Context7 MCP Pro at $10 per seat per month would spend $6,000 per year on subscriptions alone [50 × $10 × 12] (The AI Loadout). MCP gateway list prices on Oct 6, 2026 include TrueFoundry Pro at $25 per user per month (Gatana), Composio Pro at $29 per month plus $0.0003 per tool call above included usage (Gatana), and Gatana Pro at $125 per month per organization (Gatana). Docker MCP Gateway, Microsoft MCP Gateway, and IBM ContextForge are open source (MIT/Apache-2.0) with no license fee — you pay for compute, storage, and operations time.
The pricing models vary enough to change your architecture decisions. Per-user pricing (TrueFoundry $25/user/mo Gatana)) grows with headcount; per-organization (Gatana $125/org/mo Gatana)) caps the bill but examine the limits that come with the flat price; per-usage (Composio $0.0003 per tool call above credits) means agents making many calls can spike costs during high-usage periods. Every vendor also has an enterprise tier with custom pricing where SSO, SCIM, and audit logs often live. Safeguard’s model stands out: unlimited MCP tool calls on every plan including Free, with pay-as-you-go API pricing after 10,000 included deterministic calls — the argument is that metered tool calls turn security questions into budget decisions, and the answer is usually to ask less often (Safeguard).
What Should Your Audit Process Look Like?
Start by inventorying every claude mcp add line across your organization — treat each like a production service-account deployment because MCP configs hold tokens with broad access and execute outside the agent sandbox on machines that store SSH keys and cloud credentials. Run a phantom-dependency check against npm and PyPI registries for every package name in your MCP launch commands; the same category of check you already run for hallucinated dependencies in package.json (slopsquatting) needs to point at MCP server commands. Check for deprecated or archived repositories — 17.3 million downloads of deprecated packages since July 2025 have received no patches (Help Net Security).
Verify OAuth scopes and prefer read-only modes; Supabase MCP offers read-only mode, and the GitHub MCP server provides a read-only URL. Gate at the pull request level, not just the IDE — instruction files such as CLAUDE.md and AGENTS.md are context to the model, not enforced configuration, according to the vendors’ own docs. Pair a static scanner (Snyk Agent Scan, Cisco MCP Scanner) with a runtime guard (Docker MCP Gateway, Lasso, Trail of Bits mcp-context-protector) and least-privilege access. JFrog research in Australia found that 48% of organisations take a week or more to produce audit proof for a single application — your audit process should produce evidence a human can verify in minutes, not weeks.
The MCP TypeScript and Python SDKs have each passed 1 billion downloads (Hokai), and the supply of servers grew just as fast. You’re not choosing whether to govern this surface — you’re choosing whether to do it before or after an incident forces your hand. What’s the first MCP server entry you’d pull from a developer’s config today?
Recommended Reading
-
Stripe MCP Server: Agent Permission Boundary for FinOps
Stripe's official MCP server enables AI agents to interact with payment APIs, but its default restricted key security model is insufficient to prevent costly misuse. Production deployments require additional rate limiting, approval gates, and data governance to avoid financial loss and compliance violations.
-
46% of Teams Hit by AI Agent Incidents: MCP Audit Checklist
Shadow MCP is the critical blind spot your gateway misses. Gateways govern only their own traffic while unapproved MCP servers run as local JSON files on laptops with live credentials and no audit trail.
-
Strict JSON Schema Validation Stops MCP Prompt Injection
Over 200,000 MCP servers are exposed to prompt injection due to missing JSON schema validation, per recent security audits. Unconstrained tool parameters and outputs create universal attack surfaces that let attackers hijack AI agent workflows. This guide outlines critical validation steps to harden MCP deployments against these threats.