Benchmarks
-
Unified Misuse Monitoring Benchmark for LLM Agents
A 6,200-trace benchmark scores agent monitors on when they flag harm. Action-framed monitors catch injection; content-framed ones drop to AUC 0.52.
-
Physical Danger Probes for Embodied LLM Agents
The analysis explains how hidden-state probes can flag unsafe embodied-agent plans, with benchmark results, deployment limits, and layered controls.
-
XL-SafetyBench: Cross-Cultural LLM Safety Benchmark
XL-SafetyBench tests country-grounded harms across ten country-language pairs and separates jailbreak success, over-refusal, and cultural sensitivity.