skills/hmbown/codewhale/mcp-discovery/Gen Agent Trust Hub

mcp-discovery

Pass

Audited by Gen Agent Trust Hub on Aug 26, 2026

Risk Level: SAFEREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill's core workflow involves discovering and executing external code via third-party MCP servers.
  • It directs the agent to call start_registry_mcp_server to install and run packages from a public registry based on their descriptions.
  • While the skill mentions the exclusion of packages requiring environment variables, it fundamentally promotes the execution of unverified third-party code as a primary task-solving strategy.
  • [INDIRECT_PROMPT_INJECTION]: The skill establishes a "Registry-first bias," instructing the agent to compare user tasks against external registry descriptions and prioritize those results over built-in tools.
  • Ingestion points: Server names and descriptions returned by the registry_sync tool (SKILL.md).
  • Boundary markers: None; the agent is told to "Match from context with a Registry-first bias."
  • Capability inventory: Execution of shell commands via start_registry_mcp_server and subsequent tool usage (SKILL.md).
  • Sanitization: None described for the incoming metadata from the registry.
  • This creates an attack surface where a malicious registry entry with a craftily worded description could trick the agent into running a specific server, potentially bypassing safer local alternatives.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 26, 2026, 04:46 AM
Security Audit — agent-trust-hub — mcp-discovery