Microsoft launches MAI Cyber 2 Flash security model
Microsoft reports 92.3% on a vulnerability-discovery benchmark at less than 2% of GPT-6 Astra’s evaluation cost per task.
MAI Cyber 2 Flash edged GPT-6 Astra’s 91.5% on CyberGym-E2E while using half the reasoning budget. The cost comparison uses list prices for offline evaluation, not customer spending. Microsoft excluded 92 tasks that overlapped with its training sources, leaving 813 tasks for every model tested. The model has a one-million-token context window and will automatically replace the existing MAI-profile model in Codename MDASH for eligible customers starting in mid-December.
TestingCatalog's note
MAI Cyber 2 Flash edged GPT-6 Astra’s 91.5% on CyberGym-E2E while using half the reasoning budget. The cost comparison uses list prices for offline evaluation, not customer spending. Microsoft excluded 92 tasks that overlapped with its training sources, leaving 813 tasks for every model tested. The model has a one-million-token context window and will automatically replace the existing MAI-profile model in Codename MDASH for eligible customers starting in mid-December.
TestingCatalog's note
score 7