mcp-architecture

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

• Third-party content exposure detected (high risk: 0.75). The skill’s REQUIRED workflow is to configure/use MCP servers (e.g., “Official Servers” like GitHub/Slack/Puppeteer) where the MCP host can fetch and pass outsider-authored free text (issue/PR bodies, scraped web page text, Slack messages) into the agent’s LLM context via MCP resources/tools at runtime.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill includes commands that launch external MCP servers at runtime (e.g., "npx -y @modelcontextprotocol/server-github", "pip install mcp-server-slack", "docker run mcp-postgres-server") which fetch and execute remote code and can supply prompts/tools to the agent — notably linked server registry https://github.com/modelcontextprotocol/servers.