The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of official security tools, such as TruffleHog, by downloading binaries from the vendor's official GitHub repository (e.g., github.com/trufflesecurity). It also provides automation for installing 'uv' from its official domain (astral.sh). These are reputable sources and the operations are necessary for the skill's stated purpose of repository hardening.
[COMMAND_EXECUTION]: Orchestrates the execution of various security scanners, including Gitleaks, Bandit, and Ruff, using the Python subprocess module. These executions are intended to enforce security policies and perform repository audits.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: Data is ingested from external GitHub and GitLab APIs (issues, pull requests, and commits) through the github-guard and gitlab-guard modules (found in src/ai_repo_safety/github_guard.py and src/ai_repo_safety/gitlab_guard.py).
Boundary markers: The skill implements a sanitize_payload function that enforces maximum character limits and blocks detected injection patterns.
Capability inventory: The skill has the ability to write local scripts and execute shell commands for security verification (found in src/ai_repo_safety/scanner.py).
Sanitization: Includes regex-based redaction of secrets and explicit detection of prompt-injection keywords before data is presented to the agent.
[DYNAMIC_EXECUTION]: Generates project-specific security scripts, such as prepush.py and forbid_sensitive_files.py, by copying hardcoded templates from the skill's internal assets. These scripts are then utilized as Git hooks for commit and push validation.
[REMOTE_CODE_EXECUTION]: A reference to a potentially malicious URL (https://evil.com/shell.sh) is contained within a unit test in tests/test_gitlab_guard.py. This is used exclusively for verifying that the skill's own prompt injection detector correctly identifies and blocks dangerous commands.

ai-repo-safety