The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it is designed to ingest and parse external SKILL.md files during auditing and ecosystem scans.
Ingestion points: The AUDIT mode (Mode 1) reads content from SKILL.md files, and the ECOSYSTEM SCAN MODE recursively searches for and processes these files.
Boundary markers: While the skill documentation mentions using isolation delimiters as a security best practice, the implementation does not explicitly define specific delimiters for its own parsing of untrusted content.
Capability inventory: The skill has filesystem search capabilities (find) and generates markdown reports based on analyzed content.
Sanitization: The skill instructions include directives to sanitize or reject dynamic inputs in its GUARD mode (Mode 2).
[COMMAND_EXECUTION]: The skill includes an extensive library of destructive and malicious shell commands used as a blacklist for scanning other tools.
Evidence: File references/command-blacklist.md contains patterns for fork bombs (:(){ :|:& };:), mass filesystem deletion (rm -rf /), disk overwriting (dd if=/dev/zero), and reverse shells (nc -e /bin/sh).
Context: These commands are explicitly documented as 'STRICT REJECT' or 'HIGH RISK' patterns for the auditor to detect and block; they are not used for the skill's own operational logic.

security-auditor