MAI Cyber 2 Flash

MAI Cyber 2 Flash is Microsoft AI’s security model for agentic defence: finding vulnerabilities, recommending fixes and investigating threats across large codebases. Announced at Microsoft Ignite on 2026-11-17. It is built on Microsoft’s proprietary security data and is used only inside Microsoft’s security products. Part of the MAI series.

Microsoft’s claims

  • CyberGym-E2E (real-world vulnerability discovery): 92.3% for MAI Cyber 2 Flash against 91.5% for GPT-6 Astra, at under 2% of the cost per task, using half the reasoning budget.
  • WebExploitBench: 40.4% against 38.4% for GPT-5.6 Cyber, at the same reasoning budget.
  • Benchmark cleaning: 92 of the benchmark’s 905 tasks overlapped with training-data sources. Microsoft excluded them for all models and reported CyberGym-E2E on the remaining 813 tasks.
  • Context window: one million tokens, four times that of MAI Cyber 1.1 Flash.
  • Early evaluations on investigation-agent and security-detection-authoring tasks show performance “comparable to leading models”. No figures are given.
  • Customer report: Insight Enterprises reports 20% better token efficiency than the GPT models it tested (GPT-5.4, GPT-5.3 Codex and GPT-5.4 Mini).

Availability and access

  • Available exclusively in Microsoft security products, including Codename MDASH.
  • Eligible customers receive it as a replacement for the existing model in the MAI profile from mid-December.
  • A comparison with GPT-6 Cyber is promised “in the coming weeks”; that model is limited to select customers.

Billing

Billed per token at list prices that the post gives in a footnote. Those prices are not recorded here; check the post or Microsoft’s pricing page.

Source limits

  • Microsoft says results “reflect evaluations on specific security tasks and configurations” and vary with workload, configuration, reasoning effort and method.
  • The comparison with GPT-6 Cyber is not yet published.
  • Early-evaluation and customer results are not measured figures.

Sources