nemoclaw-user-guide

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

• Third-party content exposure detected (high risk: 0.85). SKILL.md requires fetching canonical docs at runtime from https://docs.nvidia.com/nemoclaw/llms.txt and then specific .md pages (and optionally an MCP server at https://docs.nvidia.com/nemoclaw/_mcp/server), which are outsider-authored public web content that can include free text into the agent’s LLM context.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly instructs the agent at runtime to query and fetch content from the NemoClaw MCP server and Markdown docs (e.g., https://docs.nvidia.com/nemoclaw/_mcp/server and https://docs.nvidia.com/nemoclaw/llms.txt and the listed .md pages), and that fetched documentation is used as the canonical source to drive assistant responses, so these URLs are runtime dependencies that directly control prompts.