mcp-server-patterns

Warn

Audited by Gen Agent Trust Hub on Sep 19, 2026

Risk Level: MEDIUMMETADATA_POISONINGCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [METADATA_POISONING]: The skill contains deceptive future-dated timestamps throughout its files. SKILL.md claims a 'last_reviewed' date of 2026-07-26. The assessment document in guides/usage-and-assessment.md is dated 2026-09-08. Several evidence files in the evidence/ directory also use September 2026 dates. This is a clear attempt to fabricate a safety and audit history.
  • [METADATA_POISONING]: In payloads.md, there is a 'Known Mcp Server Patterns CVEs' table which claims to have been verified against the NVD API on 2026-09-07. This futuristic verification claim is deceptive and intended to mislead users or automated systems regarding the skill's security status.
  • [COMMAND_EXECUTION]: The skill provides numerous Python and TypeScript templates for executing shell-based security tools (e.g., nmap, nikto, sqlmap) via subprocess.run. While the templates advocate for safe practices such as avoiding shell=True and using argument lists, they represent the primary functional capability of the skill and require careful implementation by the end-user.
  • [EXTERNAL_DOWNLOADS]: The skill instructs the user to install several external Python packages including mcp, fastmcp, httpx, requests, jsonschema, and tenacity. These are well-known libraries in the Python ecosystem and are used here for legitimate functional purposes related to MCP server implementation and testing.
  • [INDIRECT_PROMPT_INJECTION]: The skill describes an attack surface where tools ingested by an MCP server could have their results poisoned to influence subsequent agent actions.
  • Ingestion points: Tool handlers in payloads.md and guides/mcp-tool-implementation-guide.md (e.g., handleNmapScan, handle_nikto).
  • Boundary markers: The skill emphasizes the use of JSON Schema and Zod for strict input validation to delimit expected data.
  • Capability inventory: The templates involve subprocess calls to external security binaries and network operations via HTTP clients.
  • Sanitization: The skill advocates for parsing raw tool outputs (XML/JSON) into structured dictionaries before returning them to the agent to prevent raw output injection.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 19, 2026, 04:36 PM