Your AI Agent Passed the Approval Check. Did the Side Effect?

Описание и применение

Система Agent Action Guard обнаруживает обходы проверок разрешений у агентов OpenAI, используя статический анализ и доказательства выполнения, чтобы гарантировать, что реальные последствия защищены.

Задача

AI agents may bypass approval boundaries via alternate execution paths, leading to unintended real-world consequences.

Результат

The combined static and runtime approach successfully identified bypass paths and confirmed boundary enforcement, enabling developers to validate safety boundaries in CI pipelines.

Рабочий процесс

• 1. Define approval boundary in agent workflow.
• 2. Run static scanner to detect possible counterexample paths.
• 3. Attach Runtime Witness hooks to OpenAI Agents SDK to record tool lifecycle events.
• 4. Execute agent with hooks in guarded mode to observe boundary usage.
• 5. Analyze runtime witness output for boundary verdict.
• 6. If OBSERVED_BYPASS or UNRESOLVED_TRACE, investigate and adjust agent or approval logic.

Доказательства

The static scanner in the project looks for that second path.

Источник

The runtime layer asks: «What path was actually observed when the protected consequence occurred?»

Источник

If the execution evidence is incomplete, the system does not infer safety and does not infer a bypass.

Источник