The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes a local Python script at ~/.claude/plugins/hs/commands/hs_cmd.py to manage the plugin's status (on/off) and handle user-requested safety bypasses (/hs skip). This is a functional component of the safety tool's infrastructure.
[DATA_EXFILTRATION]: Sensitive file paths such as ~/.ssh/id_rsa, ~/.aws/credentials, and .env are listed along with exfiltration patterns like curl -d @.... Technical analysis confirms these are used as match criteria for a defensive blacklist to prevent data leaks, rather than for performing exfiltration.
[PROMPT_INJECTION]: The instructions employ strong imperative language and explicit block lists ('MANDATORY', 'NEVER', 'INSTANT BLOCK') to enforce safety boundaries. These directives are defensive constraints designed to prevent the AI from executing potentially malicious or destructive commands found in user input or external data.

hs