The Agent Skills Directory

[PROMPT_INJECTION]: The skill's reference files (references/prompt-injection-patterns.md) catalog numerous injection techniques like 'ignore previous instructions' and 'DAN mode' to provide detection context for the agent. These are intended for analysis and are not instructions for the agent to follow.\n- [REMOTE_CODE_EXECUTION]: Examples of reverse shells and dangerous command execution are provided in references/dangerous-code-patterns.md to serve as signatures for the scanner. The functional Python script (scripts/scan_skill.py) does not contain or execute these patterns.\n- [OBFUSCATION]: The skill documents and identifies techniques like zero-width characters and homoglyphs. The functional scanning logic contains regex patterns to detect these methods in target files without implementing them maliciously.\n- [DATA_EXFILTRATION]: While exfiltration patterns are described in the reference documentation for identification, the analysis script itself runs locally with bounded file access and redacts sensitive information from its reports.\n- [EXTERNAL_DOWNLOADS]: The skill correctly identifies and credits upstream helpers from Sentry's official GitHub repository, which is a well-known and reputable source for security-related patterns.

skill-scanner