setup-mcp

Warn

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructs the agent to execute a shell command python3 __SKILL_DIR__/scripts/validate_mcp.py '<json>' --merge <config_path> to update configuration files. The <json> argument is a string constructed by the agent based on user-provided input. If the agent fails to properly escape single quotes during interpolation, an attacker could potentially execute arbitrary commands by breaking out of the shell argument.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted configuration data from various external sources such as GitHub READMEs, VS Code settings, or Claude Desktop configurations.
  • Ingestion points: External MCP configuration snippets provided by the user in SKILL.md.
  • Boundary markers: The instructions explicitly tell the agent to reconstruct the JSON themselves based on their understanding of the user's intent, rather than passing the raw input directly to the validation script.
  • Capability inventory: The skill has the capability to write to the local filesystem (specifically ~/.copilot-shell/settings.json and project-local settings) and to execute shell commands via the validate_mcp.py script.
  • Sanitization: The agent is instructed to sanitize inputs by converting clear-text tokens into environment variable references ($VAR_NAME) and the Python script checks for forbidden fields like type and transport before merging.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 14, 2026, 03:58 AM