I built a scanner for AI-agent approval bypass paths

Описание и применение

Сканер обнаруживает пути обхода проверки утверждения, позволяющие ИИ-агенту развернуть код в продакшн без надлежащего одобрения.

Задача

AI agents may bypass expected approval boundaries to reach production, creating security gaps.

Результат

The scanner successfully identified counterexamples where AI agents could reach production deployment without passing through the approval boundary, providing actionable evidence for remediation.

Рабочий процесс

• 1. Install the scanner via pip. 2. Run the scanner against a target repository, specifying the consequence (e.g., production_deploy) and the boundary (e.g., approval).
• 3. The scanner builds a modeled consequence path across Python and GitHub Actions.
• 4. It searches for alternative modeled routes to the same consequence that do not cross the expected approval boundary.
• 5. If such a path is found, it outputs COUNTEREXAMPLE_FOUND, the path, and source‑code evidence; otherwise it returns UNKNOWN or COVERED_WITHIN_MODEL.

Доказательства

I built an early open-source scanner to look for this specific class of problem.

Источник

It deliberately does not claim full repository completeness.

Источник