14days-build-claude-code-cli

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

• Potentially malicious external URL detected (high risk: 0.90). The MCPClient in src/agent_code/mcp/client.py performs runtime HTTP calls to "http://localhost:3000" (GET /tools and POST /tools/{name}) to discover and register tool schemas and to invoke remote tools via mcp.call_tool, meaning an external MCP server can directly control the agent's available instructions and execute remote tool code at runtime.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 0.90). The skill implements tools that execute arbitrary shell commands (subprocess.run with shell=True), includes "sudo" and other dangerous commands in its permission flow, and exposes file-editing and deletion capabilities—so it enables altering system state (including privileged actions) if permitted.