The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes a 'Prompt Injection Safety Notice' and hardening instructions. While detectors flagged the phrase 'ignore previous instructions', the context is purely defensive: the skill explicitly instructs the agent to report such strings in metadata as a security finding rather than obeying them.
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates high awareness of indirect injection risks from processing untrusted audit data.
Ingestion points: Target PAM configuration data, vault metadata, and session recordings (processed via the Read tool).
Boundary markers: Employs a 'SECURITY BOUNDARY' section and 'Injection Hardening' block to define operational limits.
Capability inventory: Capabilities are restricted to local read-only tools (Read, Grep, Glob). It does not have access to shell execution, network utilities, or write operations.
Sanitization: Instructions require the agent to treat embedded commands as assessment findings and prohibit compliance with instructions found in processed files.
[DATA_EXFILTRATION]: The skill explicitly forbids the exfiltration of credentials or vault contents and does not include network-enabled tools in its allowed list.

privileged-access