agent-tool-routing

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

• Third-party content exposure detected (high risk: 0.85). The skill ingests tool metadata from MCP server discovery (await c.list_tools()), where t.description is LLM-facing free text authored by an external MCP server (outsider) and is then embedded/placed into the LLM tool list context.

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

• Potentially malicious external URL detected (high risk: 1.00). The skill's runtime discovery code calls Client("http://billing-mcp:8080") to fetch tool descriptions that are injected into the registry and presented to the LLM, so content from http://billing-mcp:8080 directly controls agent prompts/tool selection.