{"data":{"id":"9a545fa2-9065-4897-93ca-1ba180a77c93","title":"Nuclear-Sabotage Malware Benchmark Trips Up Most Frontier AI Models","summary":"SentinelOne created a benchmark test using the Fast16 malware (a 2005 Windows program designed to sabotage Iran's nuclear weapons development) to evaluate how well frontier AI models can conduct long-horizon reverse-engineering, which is the process of analyzing software to understand how it works. GPT-5.6 Sol was the only model tested that completed all eight stages of the investigation, while other models like GPT-5.5, GLM-5.2, and Anthropic's Opus struggled with what researchers call \"project-scale recovery,\" or the ability to fix errors and trace their consequences throughout an investigation. The researchers concluded that human oversight remains essential because even the best-performing AI made technical mistakes and needed human analysts to validate conclusions.","solution":"According to SentinelLabs researchers, \"the best current use [of these AI models] is supervised investigative agency, with human analysts defining objectives, exposing blind spots, and retaining final publication authority.\" The source emphasizes that \"Senior reverse engineers remain essential\" to oversee AI-assisted investigations.","labels":["research","safety"],"sourceUrl":"https://www.securityweek.com/nuclear-sabotage-malware-benchmark-trips-up-most-frontier-ai-models/","publishedAt":"2026-07-23T12:42:12.000Z","cveId":null,"cweIds":null,"cvssScore":null,"cvssSeverity":null,"severity":"info","attackType":[],"issueType":"news","affectedPackages":null,"affectedVendors":["OpenAI","Anthropic"],"affectedVendorsRaw":["OpenAI","GPT-5.5","GPT-5.6 Sol","Z.ai","GLM-5.2","Anthropic","Opus 4.x"],"classifierModel":"claude-haiku-4-5-20251001","classifierPromptVersion":"v3","cvssVector":null,"attackVector":null,"attackComplexity":null,"privilegesRequired":null,"userInteraction":null,"exploitMaturity":null,"epssScore":null,"patchAvailable":null,"disclosureDate":"2026-07-23T12:42:12.000Z","capecIds":null,"crossRefCount":0,"attackSophistication":"moderate","impactType":["integrity"],"aiComponentTargeted":"model","llmSpecific":true,"classifierConfidence":0.85,"researchCategory":null,"atlasIds":null}}