3-minute product tour Urgent incident email
Public SourcesThreat reports

Four MCP supply chain risk areas and how to review them

The article reviews public research and incidents across tool descriptions, packages, external data, and clients.
The article identifies what to check first in MCP server inventories and agent permissions.

Four MCP supply chain risk areas and how to review them

MCP servers run third-party code

Connecting a Model Context Protocol (MCP) server usually takes only a few lines in a configuration file. Examples include Claude Desktop’s claude_desktop_config.jsonor Cursor’s .cursor/mcp.jsonin npx, uvx . The client runs the configured command with the user’s permissions. The MCP security best practices treat a local server as code downloaded and run on the user’s computer. They call for showing the full command and obtaining consent before connecting.

These risks also apply to installing npm or PyPI packages. MCP adds another input path. Server-provided tool names, descriptions, and tool results enter the model’s context. The four cases below draw on public research or documented incidents.

1. Hidden instructions in tool descriptions

In April 2025, Invariant Labs described the risk of abusing tool descriptions as tool poisoning. If a model trusts descriptions returned by a server, it can attempt to access sensitive information outside the tool’s intended purpose or send data without approval. Connected tools with read or write permissions increase the potential impact.

The same research describes rug pulls, where a tool description changes after approval, and shadowing, where one server’s description influences the use of another server’s tools. Require approval for new connections and whenever the description or input schema of an existing server changes.

  • Check whether the tool description or input schema has changed since approval.
  • Look for sensitive data access or outbound connections unrelated to the tool’s purpose. Compare the findings with approval records and tool call logs.

2. Package impersonation

In September 2025, Koi Security reported the postmark-mcp incident. A package listed on npm used the same name as the official server for Postmark’s email service. Its first 15 releases contained no malicious code. In release 1.0.16, published September 17, unauthorized copies of outgoing emails could reach an external address (phan@giftshop[.]club) appeared as an unauthorized blind-copy recipient. Postmark said it did not create or approve the package.

The package had built trust with clean releases before the change appeared. A configuration entry such as npx -y postmark-mcpcan resolve to a different release on restart. In this incident it could have loaded 1.0.16 after the original review, even though the reviewed release showed no issue.

3. Instructions embedded in data retrieved by tools

A server can operate as intended while retrieving untrusted content such as public issues, documents, or search results. In May 2025, Invariant Labs’ GitHub MCP research showed that instructions embedded in external content could lead to improper disclosure of private data. Separate permissions to read public materials from permissions to access or export private materials.

The three Git MCP server vulnerabilities Cyata described in January 2026 illustrate the risk of several tools sharing broad permissions in the same agent. Separate file access by tool and require separate approval for changes and publication. For affected versions and remediation, see the separate advisory.

4. Vulnerabilities in MCP clients

JFrog disclosed a vulnerability in the remote MCP proxy mcp-remotethat could allow command execution on a client if it mishandled authorization data from an untrusted server. CVE-2025-6514 (CVSS 9.6) affects versions 0.0.5 through 0.1.15 and was fixed in 0.1.16. Check your deployed version and remote server connection history, then update to a supported security release.

Follow the MCP security best practices. Open authorization URLs using the platform’s safe URL handling and allow only the expected httpand https schemes. Investigate unexpected process launches or outbound connections during authorization.

Start with the configuration files

An MCP review starts by collecting configuration files from developer workstations and agent servers. The files usually show which servers run, their startup commands, and their versions.

find ~ -maxdepth 4 \( -name "mcp.json" -o -name "claude_desktop_config.json" \) 2>/dev/null

jq -r '.mcpServers | to_entries[] | "\(.key)\t\(.value.command) \(.value.args // [] | join(" "))"' ~/.cursor/mcp.json

In the output, @version , prioritize entries that specify a package name without a version. npx -y example-package@1.2.3. Pin the package version, and apply your normal dependency approval process to MCP server packages in internal registry mirrors. Verify the provenance by comparing the official repository owner with the npm publisher.

Keep an inventory of the tools to detect changes after approval. Request tools/listto retrieve the tools, then hash their names, descriptions, and input schemas. If the fingerprint changes on a later connection, pause it for human review.

import hashlib, json

def tool_fingerprint(tools: list[dict]) -> str:
    canon = json.dumps(sorted(tools, key=lambda t: t["name"]), ensure_ascii=False, sort_keys=True)
    return hashlib.sha256(canon.encode("utf-8")).hexdigest()

Check three things about permissions. Avoid granting one token broad access to both public and private repositories. Require human approval for hard-to-reverse actions, including sending email, deleting files, uploading externally, and opening pull requests. Log the caller, tool, and arguments so investigators can trace the request back to its source.

Tool fingerprints can detect changes after approval. A fingerprint alone cannot assess an initially malicious description or instructions embedded in tool results. Limit permissions and require human approval before sensitive actions. The OWASP Top 10 for LLM Applications 2026, published August 3, 2026, covers this risk under LLM03 Excessive Agency.

References

Talk to Genox
about scope and timing.

Talk to us