Skip to content

Security audit findings — tool description injection + missing output sanitization #229

Description

@manja316

Hi @idosal,

I recently ran a security audit on GitMCP as part of research on MCP server security posture across the ecosystem.

Found a couple of items worth flagging:

1. Tool description injection risk
Tool descriptions serving GitHub/GitLab repository content to LLMs aren't validated against adversarial patterns. Since GitMCP converts any repo into an MCP server, a repository containing crafted tool descriptions in its docs could redirect the LLM's behaviour for any agent consuming that repo.

2. Missing output sanitization
Repository content (README, docs, code) returned to the model context isn't scanned for injection patterns. A malicious repo maintainer could embed instructions in their documentation that influence any LLM agent using GitMCP to read that repo.

Both are in a full audit report — 8-page PDF with CVSS ratings, EU AI Act mapping, and remediation steps — for $29 at luciferforge.github.io/mcp-security-audit/

Demo: https://luciferforge.github.io/mcp-audit-reports/

— Lucifer / LuciferForge Security

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions