The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it interpolates unsanitized user text directly into prompts for sub-agents (ClaudeResearcher and GeminiResearcher). \n
Ingestion points: User-supplied text or concepts in Workflows/Explain.md. \n
Boundary markers: Absent; user content is placed directly within instructions without delimiters or safety warnings. \n
Capability inventory: Local script execution (bash), file system modification (mkdir, rm), and local network access (curl). \n
Sanitization: None; user input is not validated or escaped before interpolation. \n- [COMMAND_EXECUTION]: The skill executes a local bash script (RenderMermaid.sh) and uses system commands like mkdir and rm to manage visual assets and session storage in the ~/.claude/ directory. \n- [DATA_EXFILTRATION]: The skill performs local network requests using curl to http://localhost:8888/notify for status updates. \n- [EXTERNAL_DOWNLOADS]: The skill provides instructions for the user to download and install the @mermaid-js/mermaid-cli package from the NPM registry.

naval-explain