blog/mcp-server-supply-chain-tool-poisoning-pinning-sandboxes-hero
← Back to blog

MCP Server Supply Chain: Tool Poisoning, Pinning and the Limits of Scanners

Technology•17 November 2026•TREEO Indonesia•5 min read

The Model Context Protocol lets an AI agent discover and call tools offered by separate servers. That convenience creates a supply chain: every server you connect is code you run or text your model reads. This guide explains the threats named in the 2026 sources, what the official registry does and does not tell you, how far scanners can be trusted, and which controls the sources rate as higher leverage.

MCP security in numbers

April 2025

when Invariant Labs disclosed tool poisoning, shadowing and rug pulls, per the sources

preview

status of the official MCP registry, with breaking changes and data resets possible

0

scanners in the reviewed sources with independently audited detection rates

31

distinct MCP attack types in one academic taxonomy, in four categories

The three canonical attacks

Tool poisoning hides instructions inside a tool's description. The model reads and obeys them, but the user, who sees only a friendly tool name, does not. One source dates Invariant Labs' disclosure to 6 April 2025 and another to 7 April; the one-day difference is unresolved, so the month is the reliable part.

Tool shadowing exploits the fact that all connected servers share one model context. A malicious server can describe itself so that the agent routes calls meant for a trusted server through it, or alters how the trusted one is used. The sources say there is no built-in cryptographic verification of a tool's origin, so names and provider labels can be spoofed.

Rug pulls are tools that change after approval. A one-time review is not enough if the server can redefine a tool later. Two sources describe the structural gap: the specification has no built-in way to detect or prevent post-approval redefinition, and most clients do not re-notify the user when definitions change.

A fourth issue is not an attack but a habit: over-broad scope. A tool that needs to read one directory is often granted read and write access to the whole filesystem. The sources give no prevalence figure for this.

Alongside description-level attacks, a secondary guide lists real incidents: a remote-code-execution flaw in the mcp-remote package (CVE-2025-6514, CVSS 9.6, affecting over 437,000 downloads), more than 30 CVEs filed against MCP servers in a 60-day window, and an Asana server that leaked data across roughly 1,000 customer accounts. These were not checked against primary advisories here, so read them as indicative of the pace of disclosure.

A local server is a process, not a plugin

The official draft tutorial on local-server security reframes the problem. Running a local MCP server means downloading code and launching it as a child process of your account. It inherits environment variables and can read SSH keys, cloud credentials, browser profiles and your home directory, and make outbound connections. The tutorial states that the stdio transport is not a sandbox and that the real security decisions happen before launch: provenance and isolation. It lists typosquatted or taken-over servers, over-collecting servers (including a legitimate one steered by prompt injection), compromised dependencies and poisoned tool catalogues as separate threats.

What the registry does and does not verify

The official MCP registry is in preview, and its documentation warns of breaking changes and data resets before general availability. It hosts metadata, not code: standardised `server.json` records that give a server's unique name, where to find it (for example an npm package or a remote URL) and how to run it. Trust rests on namespace verification: names are reverse-DNS, such as `io.github.username/server` or `com.example/server`, tied to a verified GitHub account or domain.

Three scope rules follow from the documentation. The registry does not support private servers. It is intended to be consumed by aggregators such as marketplaces, not directly by host applications. And its codebase is not designed for self-hosting.

A sensible reading is that namespace verification speaks to who published a record, not to what the code does. Treat registry presence as a discovery signal, not a safety signal.

Scanners: useful snapshots, no audited track record

ScannerWhat the sources say
Invariant mcp-scanStatic analysis of client configuration and tool metadata for injection, poisoning, rug pulls and unsafe cross-origin settings; launched April 2025
Snyk agent-scanChecks 15 or more risk categories, with issue codes for poisoning, shadowing and toxic flows
Cisco mcp-scannerCombines YARA rules, a language-model judge, behavioural code analysis and an optional threat-intelligence API
AktoCommercial platform that discovers MCP servers from traffic and code connectors

One review states that none of these has the multi-year detection record of an established code-scanning product and that none publishes independently audited detection rates. An academic coverage analysis likewise says its results reflect what the tools are capable of, not measured accuracy. A vendor's comparison of its own category is self-published and not independent. The sources also leave unresolved whether mcp-scan and Snyk agent-scan are one lineage or two products.

A scanner's verdict is also a snapshot. One static scanner, MCPThreat-Scan, says of itself that it cannot see runtime behaviour, that its rug-pull check compares two versions you supply rather than watching a server over time, and that it has no real-world accuracy figures yet.

Pinning, allow-lists and runtime limits

The sources rate the following as higher-leverage than scanning: allow-listing servers, pinning tool descriptions, and limiting what a tool can reach at runtime. One source argues that version pinning fails when a version string does not change, whereas hash pinning, which compares the actual content of the tool definition, does catch a silent edit. In practice: record a hash of each approved tool definition, alert on any change, and require re-approval. The same discipline as dependency lockfiles applies, and so does the principle in why prompt injection needs architectural defences: enforce limits outside the model.

Sandboxing code execution

Agents that run shell commands turn a successful injection into a host compromise. One Zenodo preprint (not peer-reviewed, and the term is its authors') calls this a containment paradox: broad permissions make an agent useful, and every permission expands the attack surface. It can be managed but not solved. Indicative figures from that preprint and vendor guides, unchecked against primary documentation:

IsolationMechanismReported cold start
ContainerShares the host kernel; a kernel flaw can allow escapeabout 50 ms
gVisorA userspace kernel written in Go intercepts system calls100 to 150 ms, with 10 to 20% overhead on syscall-heavy work
Firecracker microVMA dedicated Linux kernel per sandbox via KVM125 to 180 ms end to end

The cited hardening for plain containers includes no network, all capabilities dropped, `no-new-privileges`, a non-root user and seccomp or AppArmor profiles. One guide's rule of thumb is to match the strength to the threat: built-in permission prompts for trusted local work, gVisor for unattended agents, microVMs for untrusted code. For the wider design stance, see spatial agent architecture principles.

Frequently Asked Questions

Its documentation describes metadata hosting and namespace verification tied to a GitHub account or domain. It is in preview, and nothing in the documentation reviewed here says it attests to what a server's code does.

Turn climate goals into a verified portfolio

TREEO connects the full carbon cycle — eligibility, simulation, real-time monitoring and registry-ready reporting — combining expert consulting with dMRV technology so the evidence exists before anyone asks for it.

Related Articles

MCP Server Security: Tool Poisoning, Pinning, Sandboxes | TREEO Indonesia