The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions and SKILL.md documentation call for the use of the Bash tool to run a bundled Python script (scripts/scan_skill.py). This tool usage is justified as the primary mechanism for the skill to perform static analysis on target directories.
[PROMPT_INJECTION]: The file references/prompt-injection-patterns.md and the scanner script scripts/scan_skill.py contain a comprehensive list of prompt injection and jailbreak patterns (e.g., 'ignore previous instructions', 'DAN mode'). These are clearly defined as signatures for a detection engine and are not intended to be executed as instructions against the host agent.
[REMOTE_CODE_EXECUTION]: The references/dangerous-code-patterns.md file includes examples of reverse shells and malicious code execution (e.g., os.system('nc -e /bin/sh ...')). These are provided as educational material and search patterns for the scanner, serving as diagnostic data rather than functional code.
[EXTERNAL_DOWNLOADS]: The skill script identifies various untrusted domains (e.g., evil.com) in its pattern matching logic to detect data exfiltration. The skill itself does not perform unauthorized external downloads; its only external dependency is the well-known pyyaml package, and it references trusted documentation domains such as Sentry's official sites and GitHub.
[SAFE]: The skill follows security best practices by providing a framework for least-privilege assessment and human-in-the-loop verification of findings. All 'malicious' indicators detected by static analysis are confirmed to be data/signatures rather than active attack vectors.

skill-scanner