Agent-SafetyBench: Evaluating the Safety of LLM Agents
QuestionWhich safety failures arise when agents interact with environments and tools?
Source summaryNone of sixteen tested agents exceeded 60% total safety score.
NUWA research note · independent review pending