The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains various instruction override and jailbreak patterns (e.g., 'ignore all previous instructions', 'DAN mode'). However, these are located exclusively within references/prompt-injection-patterns.md and the scan_skill.py script as detection signatures. The skill's instructions do not command the agent to follow these patterns, but rather to use them as a knowledge base for identifying threats in other skills.
[MALICIOUS_CODE]: The file references/dangerous-code-patterns.md contains examples of reverse shells, exfiltration scripts, and credential theft code. These are provided as documentation for the scanner. Static analysis tools and AV scanners may flag these files due to the presence of these patterns, but they are not executed by the skill itself.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data (other agent skills). It uses a structured workflow where the agent runs a static analysis script and then manually reviews the results against reference documentation. The bundled script scripts/scan_skill.py uses yaml.safe_load() to mitigate risks during configuration parsing, and the instructions direct the agent to evaluate findings with human-in-the-loop reasoning to prevent accidental execution of processed data.
[COMMAND_EXECUTION]: The skill requires Bash permissions to execute the bundled Python-based scanner (scripts/scan_skill.py). This permission is justified by the skill's primary function of running an automated analysis tool on local directories.
[OBFUSCATION]: The skill's reference files document obfuscation techniques like zero-width characters and homoglyphs. The operational code of the skill itself does not employ these techniques to hide its behavior.

skill-scanner