What changed
GuardianAgentBench tested 580 scenarios, 81 tools and five adversarial modes. Its best configuration reached a reported 74.8% accuracy, while execution-time guardrails recovered 19.9% of failures at a 0.5% false-positive rate.
Why it matters
Stronger models still omit required tools; weaker models select or repeat bad tools. More capability does not remove operational failure.
What to do
Validate arguments, required-tool coverage, duplicate calls, cost limits, authority and reversibility immediately before execution.
Caveat
This is a new preprint using many synthetic tasks and limited human validation.