Newshot

Microsoft's Most Cost-Effective Cybersecurity AI Model, MAI Cyber 2 Flash, Debuts: Performance Rivals GPT-6 Astra at About 2% of Its Cost per Task

AI summary

Microsoft's AI division released MAI Cyber 2 Flash on November 17, a specialized security model positioned as a high-cost-performance cybersecurity AI for enterprise security teams, designed to help them discover software vulnerabilities around the clock, recommend fixes, and investigate cyber threats while significantly reducing the compute and cost of using AI at scale. On the CyberGym-E2E benchmark measuring real-world vulnerability discovery, it achieved a 92.3% task success rate, slightly above GPT-6 Astra's 91.5%, at a disclosed cost per task of less than 2% of the latter's, based on the remaining 813 tasks after excluding 92 that overlapped with training data sources; on WebExploitBench it scored 40.4% under the same inference budget versus GPT-5.6 Cyber's 38.4%, and it offers a 1 million token context window, four times that of the previous-generation MAI Cyber 1.1 Flash. MAI Cyber 2 Flash will be exclusive to Microsoft security products, including the product codenamed MDASH, and from mid-December the existing models in eligible customers' MAI profiles will be automatically replaced with MAI Cyber 2 Flash.

Translation(model translation, not the source's own text)

Microsoft's Most Cost-Effective Cybersecurity AI Model, MAI Cyber 2 Flash, Debuts: Performance on Par with GPT-6 Astra at About 2% of Its Per-Task Cost

IT之家 November 18 news: Microsoft's AI division released a new purpose-built security model, MAI Cyber 2 Flash, on November 17, positioned as a "high cost-effectiveness" cybersecurity AI for enterprise security teams. The model is designed to help security personnel discover software vulnerabilities around the clock, recommend remediation plans, and investigate cyber threats, while significantly reducing the compute and cost of using AI at scale.

Microsoft says MAI Cyber 2 Flash was built on its proprietary security data and security research experience accumulated over years. On the CyberGym-E2E benchmark, which measures real-world vulnerability discovery capability, the model achieved a 92.3% task success rate, slightly above GPT-6 Astra's 91.5%, while Microsoft disclosed a per-task cost of less than 2% of the latter's. The result was calculated based on the remaining 813 tasks after excluding 92 tasks that overlapped with training data sources. IT之家 attaches the relevant screenshot below:

In another test, WebExploitBench, MAI Cyber 2 Flash scored 40.4% under the same inference budget, higher than GPT-5.6 Cyber's 38.4%. Microsoft also said the model has a 1-million-token context window, four times that of the previous-generation MAI Cyber 1.1 Flash, allowing it to analyze larger codebases in a single pass and making it better suited to complex, multi-step security investigation tasks.

Initial customer feedback also shows its efficiency advantages. An evaluation by Insight Enterprises said the MAI Cyber model is 20% more token-efficient than GPT Profile, meaning it may consume fewer compute resources when completing similar security tasks. However, Microsoft also stressed that the above performance and cost comparisons will vary with specific workloads, model configurations, inference intensity, and evaluation methods, and that the cost data reflects offline evaluation costs and does not equal customers' actual spending.

In terms of deployment, MAI Cyber 2 Flash will be exclusive to Microsoft security products, including a product codenamed MDASH. Starting in mid-December, existing models in the MAI profiles of eligible customers will be automatically replaced with MAI Cyber 2 Flash.

Related Articles

  • Microsoft announces PowerShell 7 will ship with Windows by default, with Win11 26H2 the last version to pre-enable 5.1
  • "The world's No. 1 photo-editing AI model": Microsoft MAI Image 2.7 debuts, beats GPT Image 2.6 to take the top spot at 1/5 the price
  • Google updates Gemini privacy controls, allowing chat history retention without participating in AI training
  • Microsoft's most cost-effective coding AI model: MAI Code 2 reaches about 96% of Claude Opus 5.5's performance at 5% of the cost
  • Microsoft Win11 26H2 memory optimization tested: experience on 8GB machines improves significantly
  • Multiple well-known websites tampered with, attackers steal cryptocurrency and compromise iPhones

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗