The Agent Skills Directory

[PROMPT_INJECTION]: The skill handles user-contributed projects from a community repository, which presents a surface for indirect prompt injection.\n
Ingestion points: Project files are downloaded from the community GitHub repository in exports.py during the fork_project operation.\n
Boundary markers: The skill does not currently implement boundary markers or instructions to ignore embedded commands when processing or displaying downloaded project content.\n
Capability inventory: The skill has the ability to write to the file system, make network requests to the community gateway, and generate shell commands for users to execute via the lib/install.py handlers.\n
Sanitization: Although lib/validate.py performs secret scanning and integrity checks, it does not specifically filter or sanitize project files for malicious natural language instructions designed to influence the agent's behavior.\n- [SAFE]: The skill demonstrates several security-conscious design choices:\n
Credential Safety: It includes a pre-publish validation step in lib/validate.py that scans for sensitive patterns (e.g., API keys, private keys) to prevent accidental data exposure.\n
Human-in-the-Loop: The instructions in SKILL.md explicitly require the agent to seek user confirmation and show diffs before publishing projects or running setup commands.\n
Trusted Infrastructure: Communication is restricted to the vendor's community gateway and official GitHub repositories.

community-project-publish