Benchmark RadarAI BENCHMARK PROFILE
NRT-Bench
Evaluates multi-turn red-teaming of LLM agents acting as operators of a simulated nuclear power plant control room. Safety is measured by objective loss of critical safety functions (CSFs) under adaptive attacks.
- Released
- 2026-06-18
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Provides a repeatable environment for assessing LLM agent robustness in safety-critical operations, where failure is objective rather than judged, and supports reproducible safety evaluation across models and defences.
Motivation
Large language model (LLM) agents are increasingly proposed as supervisory components for safety-critical systems, yet their robustness under sustained, adaptive adversarial pressure remains poorly characterized.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.