MAI Cyber 2 Flash
Intro
MAI Cyber 2 Flash is the specialist security model Microsoft AI introduced on 17 November 2026 (Ignite · Models) for finding vulnerabilities, recommending fixes, and investigating threats. Microsoft says it is built on the company’s proprietary security data and expertise, aimed at leading real-world security performance at a fraction of the cost of larger models, and that the capability is being brought directly into Microsoft’s security products.
It is a different specialist SKU in the same Ignite wave as MAI Code 2 / Flash and the already filed MAI Image 2.7. The previous security name in this line is MAI Cyber 1.1 Flash; 2 Flash is described as the latest specialist version, with a one-million-token context window that Microsoft says is four times larger than 1.1 Flash, so it can see more of a codebase during multi-step investigations.
Availability is narrower than for the coding models. The post says it is “coming exclusively to Microsoft security products, including Codename MDASH,” and that it will automatically replace the existing model in the MAI profile for eligible customers starting mid-December. It is introduced first as a model tier inside Microsoft security products, not as a broadly listed third-party API model.
Advertisement
Explain
The headline comparisons are CyberGym-E2E (real-world vulnerability discovery) and WebExploitBench. Microsoft reports 92.3% task success for MAI Cyber 2 on CyberGym-E2E versus 91.5% for GPT-6 Astra, at less than 2% of the cost per task, and says it also outperforms GPT-5.6 Cyber on CyberGym-E2E and WebExploitBench at the same reasoning budget. Footnotes add that 2 Flash used half the reasoning budget on CyberGym-E2E; that WebExploitBench scores were 40.4% versus 38.4%; and that cost per task used all input, cached input, and output tokens at list prices (MAI Cyber 2 Flash $0.10 / $0.01 / $0.50 per 1M tokens; GPT-6 Astra on Azure OpenAI Global $10 / $1 / $50), reflecting offline evaluation cost rather than customer spend.
A benchmark-integrity note matters: while reviewing results, Microsoft found that 92 of 905 CyberGym-E2E tasks came from the same open-source projects and vulnerability reports as some of its training data, though the benchmark’s reference solutions were not used in training. Conservatively, those 92 tasks were dropped for every model, and results are reported on the remaining 813. The post also cites early Microsoft Perception evaluations as comparable to leading models on investigation-agent and security-detection authoring, and a customer report from Insight Enterprises of about 20% better token efficiency versus the GPT configurations it compared. Microsoft adds that GPT-6 Cyber is currently available only to select customers, and that agentic, tool-using comparisons still require additional safeguards, with results planned for later weeks.
Usage
According to the official getting-started note, MAI Cyber 2 Flash is coming exclusively through Microsoft security products, including Codename MDASH, and is scheduled to replace the existing model in the MAI profile for eligible customers starting mid-December. Timing, eligibility, and how the switch works follow those product docs, rather than a Copilot or Foundry model picker like Code 2.
Microsoft positions it for around-the-clock vulnerability discovery, fix recommendations, and threat investigation, with the 1M-token context intended for larger codebases and multi-step work. The list prices in the footnotes are for the official cost comparison and are not the same as a customer bill inside MDASH. Do not confuse this entry with the already separate MAI Code 2 or MAI Image 2.7 filings, and do not read the still-pending GPT-6 Cyber comparison as a completed public bake-off.