Agent security has two failure modes: miss the attack, or break the work.

A guardrail that blocks everything is safe. It is also useless. So we benchmarked both sides of AI...