native-mcp

HIGH W007: Insecure credential handling detected in skill instructions.

• Insecure credential handling detected (high risk: 0.90). The prompt includes examples that embed secret values (e.g., "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_...", "NOTION_API_KEY": "secret_...", and "API_KEY": "secret") which encourages or requires the agent to place or reproduce secrets verbatim in .mcp.json, creating an exfiltration risk.

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.65). The skill’s runtime path is that Claude Code launches configured MCP servers (e.g., via command/args or url in .mcp.json), and those servers can return arbitrary tool outputs that Claude ingests as LLM-readable text; if any configured MCP server is remote/public or processes outsider-authored content (e.g., GitHub/Notion/web pages), that content can flow into the agent context via tool results.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The skill configuration launches third-party packages at runtime (e.g., via "npx -y @modelcontextprotocol/server-filesystem", which pulls and executes code from the npm registry such as https://www.npmjs.com/package/@modelcontextprotocol/server-filesystem), so remote code is fetched and executed as a required dependency for the agent's tools.