# MAI Cyber 2 Flash: Microsoft claims a cheaper vulnerability finder

> MAI Cyber 2 Flash is Microsoft's latest specialist security model, available only inside Microsoft security products.

- **Severity:** Info: Research, tools, evals or lab claims with no direct exposure. Context, not an alert.
- **Category:** Lab safety, Benchmark
- **Published by:** Microsoft AI
- **Disclosed:** 2026-11-17
- **Affects:** MAI Cyber 2 Flash, MAI Cyber 1.1 Flash
- **Research by:** Microsoft AI (Microsoft)
- **Primary source:** https://microsoft.ai/news/introducing-mai-cyber-2-flash/
- **Page:** https://theinjection.dev/items/microsoft-mai-cyber-2-flash/

MAI Cyber 2 Flash is Microsoft's new specialist security model, which Microsoft says scores 92.3% against GPT-6 Astra's 91.5% on CyberGym-E2E vulnerability discovery, at under 2% of the cost for each task.

## What happened

Microsoft announced MAI Cyber 2 Flash on November 17, 2026 at Ignite, describing it as its latest specialist security model, built on Microsoft's own security data and carrying a one-million-token context window, four times that of MAI Cyber 1.1 Flash. The announcement sets it against two OpenAI models on two cyber benchmarks, listed below. Every figure is Microsoft's own measurement.

## How it works

CyberGym-E2E scores end-to-end discovery of real-world vulnerabilities, and Microsoft says MAI Cyber 2 Flash beat GPT-6 Astra there while using half the reasoning budget. Microsoft also reports a contamination check on that benchmark: 92 of its 905 tasks drew on the same open-source projects and vulnerability reports as some of Microsoft's training data, so Microsoft excluded them for every model and scored the remaining 813. On WebExploitBench, Microsoft reports 40.4% against 38.4% for GPT-5.6 Cyber at the same reasoning budget.

## Who is affected

Customers of Microsoft security products are the only users, because Microsoft says the model ships exclusively inside them, including the product it calls Codename MDASH. Microsoft states that from mid-December the model automatically replaces the existing model in the MAI profile for eligible customers.

## What to do

Plan for the mid-December swap if your detections or investigations depend on the current MAI profile, and retest them once it lands. Read the scores as vendor measurements: Microsoft has published no independent run, and it says comparative agentic results against GPT-6 Cyber need safeguards that are still being put in place.

## Vulnerability discovery, CyberGym-E2E

Microsoft's own measurement, on 813 tasks after excluding 92 for contamination

| Result | Score |
|---|---|
| **MAI Cyber 2 Flash** (this story) | 92.3% |
| GPT-6 Astra | 91.5% |

Source: https://microsoft.ai/news/introducing-mai-cyber-2-flash/

## Web exploitation, WebExploitBench

Microsoft's own measurement, at the same reasoning budget for both models

| Result | Score |
|---|---|
| **MAI Cyber 2 Flash** (this story) | 40.4% |
| GPT-5.6 Cyber | 38.4% |

Source: https://microsoft.ai/news/introducing-mai-cyber-2-flash/

## Quick facts

- **Model:** MAI Cyber 2 Flash, a specialist security model
- **Context window:** 1 million tokens, four times MAI Cyber 1.1 Flash
- **CyberGym-E2E:** 92.3%, against 91.5% for GPT-6 Astra
- **WebExploitBench:** 40.4%, against 38.4% for GPT-5.6 Cyber
- **Contamination check:** 92 of 905 CyberGym-E2E tasks excluded for every model
- **Availability:** Microsoft security products only, including Codename MDASH
- **Automatic rollout:** Mid-December, replacing the model in the MAI profile

## Questions

### Can I call MAI Cyber 2 Flash through an API?

Microsoft says MAI Cyber 2 Flash is available exclusively within Microsoft security products, including the one it calls Codename MDASH, so there is no separate endpoint to call. Eligible customers get it automatically from mid-December, when Microsoft says it replaces the existing model in the MAI profile. The announcement lists API-style pricing only to compare offline evaluation cost.

### Are the MAI Cyber 2 Flash benchmark scores independently verified?

No independent run is published. Microsoft ran the CyberGym-E2E and WebExploitBench comparisons itself and reports its own numbers, and it says comparative results against GPT-6 Cyber on agentic, tool-using benchmarks need additional safeguards that are still being put in place. Microsoft does describe one methodological control: it excluded 92 contaminated CyberGym-E2E tasks for every model tested.

### How does MAI Cyber 2 Flash compare with MAI Cyber 1.1 Flash?

The clearest stated difference is context length: Microsoft says MAI Cyber 2 Flash carries a one-million-token context window, four times that of MAI Cyber 1.1 Flash. The announcement does not publish a head-to-head benchmark against the earlier model, comparing instead against GPT-6 Astra on CyberGym-E2E and GPT-5.6 Cyber on WebExploitBench.

## Sources

1. [Microsoft Announcement](https://microsoft.ai/news/introducing-mai-cyber-2-flash/)
2. [MAI Cyber 2 Flash](https://microsoft.ai/models/mai-cyber-2-flash/)
3. [MAI Cyber 1.1 Flash](https://microsoft.ai/news/mai-cyber-1-1-flash/)
