learn-mcp

Pass

Audited by Gen Agent Trust Hub on Aug 30, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONREMOTE_CODE_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes local Python scripts (e.g., python3 phases/13-tools-and-protocols/06-mcp-fundamentals/code/main.py) and optional TypeScript runners to demonstrate protocol behavior.
  • [EXTERNAL_DOWNLOADS]: Fetches tutorial configuration (learning-paths/model-context-protocol.json), documentation, and quizzes from the author's repository on GitHub (rohitg00/ai-engineering-from-scratch). It also offers to clone this repository with user confirmation.
  • [DATA_EXFILTRATION]: While the skill records "wire evidence" (JSON-RPC requests/responses) to a local file MCP-LEARNING.md, it contains explicit instructions to redact tokens, secrets, cookies, authorization headers, and sensitive parameter values from all logs.
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from remote JSON and Markdown files to drive the lesson content.
  • Ingestion points: Reads learning-paths/model-context-protocol.json, docs/en.md, and quiz.json from the repository (SKILL.md).
  • Boundary markers: None identified for delimiting external content from agent instructions.
  • Capability inventory: Includes file system writes (MCP-LEARNING.md), shell command execution (python3), and network fetching via URL.
  • Sanitization: Includes mandatory redaction of credentials and sensitive values from recorded evidence.
  • [REMOTE_CODE_EXECUTION]: The skill downloads and executes scripts from the author's GitHub repository. Security is managed through user confirmation requirements before cloning and a 'Public Deployment Gate' that blocks network-facing operations until specific security lessons are completed and manual authorization is granted.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 30, 2026, 10:39 PM
Security Audit — agent-trust-hub — learn-mcp