Microsoft Releases MAI-Cyber-1-Flash, a 5B-Active Cyber Model Achieving 95.95% on CyberGym
Microsoft
Anthropic
Microsoft AI has released MAI-Cyber-1-Flash, its first cyber defense model. Integrated with the MDASH multi-agent system, it achieves 95.95% on CyberGym, outperforming competitors by roughly 12 points. The model has 137B total parameters with 5B active and handles up to 90% of MDASH tasks, reducing costs by 50%.
Microsoft AI launched MAI-Cyber-1-Flash, a cybersecurity-specialized fine-tune of MAI-Code-1-Flash, built for cyber defense. It runs inside MDASH, Microsoft's multi-model agentic scanning harness. The model is a transformer with sparse Mixture-of-Experts layers, 137B total parameters, 5B active, and a 256k context length. On CyberGym, a suite of 1,507 vulnerability reproduction tasks, MDASH with MAI-Cyber-1-Flash and GPT-5.4 scored 95.95%, roughly 12 points above Anthropic's Mythos. MDASH manages over 100 specialized agents through stages: Prepare, Scan, Validate, Dedupe, Prove. MAI-Cyber-1-Flash handles up to 90% of tasks, escalating the hardest 10% to GPT-5.4, yielding a 50% cost saving. The model scores zeros on ExploitGym by design, as it is trained for defensive patching, not exploit generation. MDASH-assisted work previously generated 16 CVEs in Windows networking. The model is available via Microsoft's security tools.
- Сокращения
- MDASH = Multi-model Agentic Scanning Harness — Мультимодельный агентный сканирующий харнес
- CVEs = Common Vulnerabilities and Exposures — Общие уязвимости и экспозиции
- MSRC = Microsoft Security Response Center — Центр безопасности Microsoft
- ASan = AddressSanitizer — Адресный санитайзер
- OSS-Fuzz = Open Source Software Fuzzing — Фаззинг открытого кода
Source: MarkTechPost —
original