The Agent Skills Directory

[COMMAND_EXECUTION]: The toolkit provides a hardened subprocess wrapper in scripts/_lib/safe_subprocess.py that enforces secure defaults. It disables shell-based command execution, provides a clean environment to prevent host secrets from being inherited by child processes, and implements robust process tree management via Windows Job Objects or POSIX sessions to ensure clean resource cleanup.
[DYNAMIC_EXECUTION]: The skill provides standard patterns and templates for security research and exploit development using industry-standard libraries such as pwntools, capstone, and unicorn. These patterns are educational in nature and strictly align with the skill's primary purpose for security engineering and automation.
[INDIRECT_PROMPT_INJECTION]: The skill contains several components designed to sanitize untrusted input before it is processed by the AI model. Utilities like redact_headers.py, secret_scan.py, and unicode_scan.py act as a data boundary, redacting sensitive authentication tokens, credentialed URLs, and invisible steering characters from network traffic and arbitrary text.
[SAFE]: The skill implements an ActionGuard system that enforces engagement scope and rules of engagement (ROE). It gates potentially destructive actions, requiring human approval for commands like recursive deletion or force-pushing, and prevents unauthorized modification of scope definitions.

coding-mastery