NRT-Bench: Frontier LLM Agents Compromised in 8.7–12.1% of Multi-Turn Attacks on a Safety-Critical Sim

arXiv·low signal

NRT-Bench (arXiv 2606.20408, submitted June 18) stress-tests five-role LLM-agent operator teams on a simulated nuclear plant, finding 8.7–12.1% of multi-turn adversarial sessions compromised at least one critical safety function across four frontier models. Vulnerabilities were nearly disjoint — an attack that beats one model often fails on another — and identical guardrails sometimes increased attack success. For agent builders the lesson is that multi-turn red-teaming surfaces failures single-turn evals miss; the authors released the environment and attack dataset.