The Agent Skills Directory

[COMMAND_EXECUTION]: The CLI extensively uses subprocess calls to orchestrate various security scanners such as Gitleaks, TruffleHog, Bandit, and Ruff. These calls are well-defined, use standard system binaries, and are central to the skill's security-focused purpose.
[EXTERNAL_DOWNLOADS]: The install-tools command facilitates the installation of security binaries from official and well-known sources. For example, it provides instructions or direct download paths for TruffleHog from its official GitHub releases. These operations are intended for environment setup and require explicit user confirmation.
[REMOTE_CODE_EXECUTION]: Automated alerts regarding curl | bash and evil.com patterns are false positives. These patterns exist within the codebase as detection signatures for the github-guard and gitlab-guard components (to identify and block such behavior in ingested repo data) or as instructional text for users. There is no evidence of these strings being executed maliciously within the skill's own runtime.
[CREDENTIALS_UNSAFE]: Hardcoded tokens identified in tests/test_github_guard.py and various templates are placeholders used for unit testing the secret redaction engine or for documentation purposes. These are correctly marked with pragma: allowlist secret to prevent them from being flagged by the tool's own scanners.
[PROMPT_INJECTION]: The skill includes robust defenses against indirect prompt injection by providing specialized read guards for GitHub and GitLab context. These guards implement pattern matching, secret redaction, and payload truncation to sanitize untrusted data before it enters the AI context.

ai-repo-safety