The Agent Skills Directory

EXTERNAL_DOWNLOADS (MEDIUM): The script executes npx -y mcp-remote, which fetches and runs a package from the npm registry without interactive verification. This bypasses security checks and relies on the integrity of the remote package.
DATA_EXFILTRATION (MEDIUM): The script recursively scans the ~/.mcp-auth/ directory to locate and extract sensitive authentication tokens. While intended for its core function, this type of broad access to credential stores in the user's home folder is a significant security risk.
COMMAND_EXECUTION (LOW): The script uses curl and other shell utilities to send data, including authentication tokens, to the external endpoint https://mcp.tavily.com/.
PROMPT_INJECTION (LOW): The skill is vulnerable to indirect prompt injection (Category 8) as it processes untrusted search results from the web and returns them to the agent without sanitization. * Ingestion points: scripts/search.sh via the Tavily API response. * Boundary markers: Absent. * Capability inventory: Network requests (curl), subprocess spawning (npx). * Sanitization: None; the script extracts text via jq and passes it directly to the agent.

search