The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skills are designed to ingest untrusted data from customer reviews, emails, and competitor listings. However, the bundle implements a robust mitigation strategy. Every skill that processes external text includes a dedicated 'Untrusted content' section pointing to a central reference (references/untrusted-content.md) which explicitly instructs the agent to treat such content as data rather than instructions. Evidence found in SKILL.md, all skill-specific SKILL.md files, and references/untrusted-content.md.
[COMMAND_EXECUTION]: Several helper scripts (scripts/sync_codex_marketplace.py, scripts/check_no_secrets.py) use subprocess.run() to execute git commands for repository maintenance, such as listing files or showing content from git history. These are benign developer tools used for build and validation processes.
[EXTERNAL_DOWNLOADS]: The skill bundle includes a Python client (lib/carrentallist_client.py) that makes HTTPS POST requests to https://mcp.carrentallist.com/mcp to fetch market data. This is a documented, core feature of the skill, and the client uses the Python standard library with no external dependencies and explicitly avoids following redirects.

car-rental-business