RefusalGuard-M: a Scalable Human–machine Framework for Multi-Turn LLM Jailbreak Evaluation Via Semantic Refusal Manifold Modeling

Abstract Existing multi-turn jailbreak evaluation methods increasingly rely on large language models (LLMs) as automated judges to reduce the cost and scalability limitations of human assessment. However, recent studies show that LLM-based evaluators can diverge from human judgments under adversarial strategies involving subtle linguistic and semantic variations, raising reliability concerns in safety-critical domains such as cybersecurity. To address this challenge, we propose Refusal Manifold Guard (RefusalGuard-M), an open-source semantic evaluation framework that constructs a semantic refusal manifold from human-validated refusal responses for assessing LLM jailbreak interactions, including multi-turn scenarios. RefusalGuard-M uses embedding-based geometric representations to measure deviations from refusal behavior, providing a lightweight, interpretable, and reproducible alternative to LLM-based judging. We evaluate the framework across AdvBench, HarmBench, and CyMulTenSet, covering diverse jailbreak strategies, linguistic transformations, and multi-turn scenarios. Results show that RefusalGuard-M achieves strong agreement with human annotations and comparable recall performance to GPT-based evaluators while adopting a conservative evaluation strategy that prioritizes the detection of harmful outputs. On CyMulTenSet, which evaluates past-tense reformulated multi-turn jailbreaks, RefusalGuard-M achieves up to 0.87 recall, compared with 0.86 for GPT-5 and 0.81 for GPT-4, and reduces inference overhead by up to 3.7 $$\times$$ × relative to embedding-based baselines. These findings demonstrate that semantic refusal representations provide an efficient and scalable approach for jailbreak evaluation, particularly in cybersecurity settings where minimizing missed harmful outputs is critical.