The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion and analysis of untrusted third-party codebases during audit activities, creating a surface where adversarial code could influence agent behavior.
Evidence: The reconnaissance and hunting phases (RECONNAISSANCE.md, HUNTING.md) involve processing external source code that could contain adversarial instructions targeting the agent.
Risk Surface: The agent analyzes untrusted data while possessing powerful capabilities such as shell command execution and file writing.
Mitigations: The skill includes rigorous multi-phase validation (VALIDATION-AND-REPORTING.md) and independent verification steps (Phase 6) to cross-check findings and prevent accidental obedience.
[COMMAND_EXECUTION]: The auditing methodology involves executing shell commands for environment setup, repository navigation, and result validation.
Evidence: Methodology requires running commands like ls, mkdir, and node to manage findings and the audit environment (audit-methodology/SKILL.md).
[DYNAMIC_EXECUTION]: The skill executes a bundled Node.js script for finding validation and supports running target code in test harnesses for dynamic verification.
Evidence: The script audit-methodology/validate-findings.cjs is executed via Node.js during the reporting phase. The methodology also describes building and running minimal harnesses for suspect code to prove exploitability (HUNTING.md).
[EXTERNAL_DOWNLOADS]: Documentation provides instructions for fetching updated security datasets from external repositories.
Evidence: data/asvs/README.md and data/fiasse/README.md provide instructions for updating reference data via git clone and curl from official GitHub repositories belonging to the OWASP foundation and established community contributors.

kernel-audit-security