Overview
MAI Cyber 2 Flash is Microsoft AI's security-specific model, released on November 17, 2026, with a deliberately narrow remit: find flaws, suggest fixes, investigate threats for enterprise security teams. It does not write copy and does not do general conversation. The measure Microsoft reports is not general intelligence but success on CyberGym-E2E, which came in at 92.3%.
Cost is the other half of the pitch. Microsoft puts the per-task cost at roughly 2% of GPT-6 Astra. Security teams are overwhelmed by alert volume rather than by a shortage of analysis tools, so a cost measured per thousand alerts says more than a rate per million tokens. At this level the model can sit on a routine scanning pipeline instead of being switched on only for incidents.
The previous generation, MAI Cyber 1 Flash, was a sparse mixture-of-experts model with 137B total parameters activating 5B per inference, driving Microsoft's internal MDASH system and scoring 95.95% on CyberGym. Cyber 2 Flash is measured on CyberGym-E2E, which tests end-to-end task completion, so the two numbers are not comparable. That distinction matters, because lined up side by side they read as a drop when they are actually different tests.
One detail in the release notes deserves attention. In an earlier round Microsoft ran a security model across six widely used open-source projects, and every report went through a human security researcher before reaching maintainers. The maintainers ultimately received 99 verified private reports and have started fixing. The process addresses a separate problem: flooding maintainers with machine-generated flaw reports buries the real ones in projects that already lack people.
Whether a team should adopt this kind of model has little to do with leaderboards. First, is there a stable verification step in the workflow? Without one, false positives become the team's problem. Second, is the job discovery, triage or remediation advice? The three demand different things from a model. Third, is cost counted per alert or per task? The two framings lead to very different conclusions.
Key Features
- 92.3% on end-to-end security tasks: CyberGym-E2E success rate of 92.3%, scoring full tasks from detection through handling rather than single-step recognition accuracy.
- About 2% of the flagship's cost per task: Microsoft puts per-task cost at roughly 2% of GPT-6 Astra. With alert volume far outstripping headcount, that price decides whether the model can live on a pipeline.
- Discovery, remediation, investigation: Built for flaw discovery, fix suggestions and threat investigation. Each demands different output from a model, and Microsoft does not collapse them into one score.
- Sparse activation, Flash tier: Continues the Flash positioning: the prior generation ran 137B total parameters while activating 5B per inference, which keeps high-frequency calls affordable.
- Human review before maintainers see it: In an earlier scan every report passed a human security researcher first, and maintainers received 99 verified private reports. The process exists to avoid burying them.
- Built for security operations: No general chat or writing. Inputs and outputs are organized around code and threat intelligence, and it plugs into existing security operations rather than replacing them.
Use Cases
- Routine code security sweeps: scheduled repository scans that amortize the 2% cost across every day
- Alert triage and severity assignment: filter with the model before a human looks, cutting repeat work
- Remediation drafts: propose patch directions for engineers to accept or reject
- Threat investigation support: stitch scattered logs and intelligence into a readable timeline
- Coordinated open-source scanning: pair with human review so maintainers do not get flooded
- Coverage for understaffed security teams: start with high-volume, lower-risk work
Pros
- 92.3% on CyberGym-E2E measures the whole workflow rather than a single detection step
- Per-task cost near 2% of GPT-6 Astra makes routine scanning economically viable
- Discovery, remediation advice and investigation are kept separate, so the scope of use is clear
- The release documents human review before delivery, with 99 verified reports reaching maintainers
- Sparse activation in the Flash tier keeps compute predictable under high call volume
- Shares lineage with Microsoft's existing security operations, so integration is short
Pricing
How it is quoted: Microsoft gives per-task cost as roughly 2% of GPT-6 Astra and does not publish a separate per-million-token input and output rate. Prior generation for reference: MAI Cyber 1 Flash was a sparse mixture-of-experts model with 137B total parameters activating 5B per inference, scoring 95.95% on CyberGym, which is a different measure from CyberGym-E2E and not directly comparable. Availability: shipped in the MAI batch announced at Microsoft Ignite 2026 for enterprise security teams, delivered through existing security operations or an Azure environment.
Summary
MAI Cyber 2 Flash is Microsoft AI's security-specific model from November 17, 2026, built for flaw discovery, remediation advice and threat investigation. It reaches 92.3% on CyberGym-E2E end-to-end tasks, and Microsoft puts per-task cost near 2% of GPT-6 Astra. The prior MAI Cyber 1 Flash ran 137B total parameters activating 5B and scored 95.95% on CyberGym, a different measure that cannot be compared directly. The release notes describe an earlier scan of six open-source projects where human researchers reviewed every report before maintainers received 99 verified private submissions. Microsoft has not published a separate per-million-token rate for this model.
Version History
- MAI Cyber 2 Flash launches (2026-11-17): Microsoft AI ships a security-specific model for flaw discovery, remediation advice and threat investigation, reaching 92.3% on CyberGym-E2E at roughly 2% of GPT-6 Astra's per-task cost. Announced in the seven-model MAI batch at Ignite 2026.