The Agent Skills Directory

[PROMPT_INJECTION]: The skill audits untrusted codebases and reads project-specific instruction files (e.g., CLAUDE.md, .cursorrules), which constitutes an indirect prompt injection surface. Malicious code or instructions within the audited repository could attempt to influence the auditing subagents. This is mitigated by a verification loop involving 'skeptic' agents and a requirement for evidence-based findings. * Ingestion points: Entire codebase, instruction files, and manifests. * Boundary markers: Finding schema requires traces and quotes; verification loop uses skeptics to validate claims. * Capability inventory: The orchestrator and subagents can execute shell commands for reconnaissance (grep, wc) and verification (tests, linters). * Sanitization: Structured reporting schema and independent verification by non-author agents (skeptics).
[COMMAND_EXECUTION]: To verify its findings, the skill automatically detects and runs project-defined 'gate' commands (such as npm test, pytest, or cargo check) found in repository manifests. This execution of code from the audited repository is a standard part of the auditing workflow but involves running code from an untrusted source.
[CREDENTIALS_UNSAFE]: The skill's security audit lens specifically searches the codebase for hardcoded secrets, API keys, tokens, and connection strings. While this is a security-enhancing feature, it involves identifying and recording sensitive data within the tool's local audit artifacts (.nuke/findings.md).
[REMOTE_CODE_EXECUTION]: During the preflight phase, the skill may attempt to install missing environment dependencies (e.g., linters or testers) to unblock the audit process. This results in the execution of package management tools to alter the local environment if it deems the environment incomplete.

nuke-audit