The Agent Skills Directory

[PROMPT_INJECTION]: The static analysis trigger regarding instruction overrides was identified as a false positive. The skill uses defensive instructions within a 'SECURITY BOUNDARY' section to teach the agent to identify and flag adversarial content in untrusted input files, rather than attempting to hijack agent behavior.
[DATA_EXFILTRATION]: Analysis of tool usage (Read, Grep, Glob) and instructions confirms that the skill is restricted to assessment activities. It includes explicit prohibitions against exfiltrating credentials or secrets found during the review.
[SAFE]: The skill exhibits an 'attack surface' for indirect prompt injection as it ingests untrusted PAM configuration data from files provided in $ARGUMENTS. However, it incorporates strong defensive markers (a 'SECURITY BOUNDARY' block), relies on a limited capability set (read-only tools), and provides explicit sanitization instructions to the agent to prevent obedience to embedded malicious instructions.

privileged-access