Nuclear-Sabotage Malware Benchmark Trips Up Most Frontier AI Models - SecurityWeek

SentinelOne, a cybersecurity company, has developed a benchmark to test the ability of frontier AI models to reverse-engineer malware. The benchmark uses the Fast16 malware case, which is designed to simulate a nuclear sabotage scenario. The company tested several leading models, including OpenAI's GPT-5.5 and GPT-5.6 Sol, Z.ai's GLM-5.2, and Anthropic's Opus 4.x.

According to the report, most of the models failed to successfully reverse-engineer the malware. The benchmark was designed to evaluate long-horizon reasoning and cybersecurity capabilities. SentinelOne's testing revealed significant gaps in the models' ability to handle complex, multi-step tasks related to critical infrastructure protection.

The results underscore ongoing challenges in applying large language models to cybersecurity tasks that require deep technical understanding and sustained attention. The benchmark is intended to help improve model performance in detecting and mitigating sophisticated threats.