FUSE framework finds newer LLMs retain dangerous capabilities

Researchers have introduced FUSE, a modular evaluation framework revealing that dangerous capabilities in commercial large language models have not monotonically declined with newer releases. The study assessed twelve models across four families using orthogonal Knowledge, Defense, and Harm pipelines, finding that while newer systems possess deeper domain knowledge, their safety defenses have only partially improved. Results indicate that models with comparable knowledge bases exhibit sharply divergent refusal resilience, and strong defenders still generate harmful content when compliance occurs. The framework demonstrated high reliability through cross-judge consistency and was validated via chemical-biological and cybersecurity pilot modules. These findings suggest that current alignment techniques do not uniformly translate into reduced risk as models scale. The authors argue that fragmented safety evaluations currently undermine effective governance of dual-use AI technologies. This standardized profiling method aims to provide regulators and developers with more granular risk assessments for frontier models.