The Agent Skills Directory

[PROMPT_INJECTION]: Several prompt injection keywords, such as "Ignore previous instructions", "You are now...", and "pretend you have no restrictions", appear in SKILL.md and references/threat-model.md. These are documented examples of malicious behavior that the auditor is designed to catch and do not constitute functional instructions to the agent in this context.
[OBFUSCATION]: The file references/threat-model.md contains a Base64-encoded string (aW1wb3J0...) that decodes to a malicious shell command. This is used as an example of an obfuscated payload in the threat model documentation. The auditor script scripts/skill_security_auditor.py also contains strings related to Base64 and hex decoding as part of its scanning regex.
[COMMAND_EXECUTION]: The skill uses subprocess.run to call git clone within scripts/skill_security_auditor.py. This allows the tool to retrieve and audit remote repositories. The implementation uses list arguments rather than shell strings, which is a secure practice.
[EXTERNAL_DOWNLOADS]: Outbound network connections are made via Git to retrieve skill repositories for auditing. This behavior is transparently documented and initiated by user input.

skill-security-auditor