Microsoft has unveiled MAI-Cyber-1-Flash, its first generative artificial intelligence model designed specifically for cybersecurity, as the company steps up efforts to strengthen its security business through AI-powered threat detection.
During an event in San Francisco, Microsoft said the model, when paired with OpenAI’s GPT-5.4, outperformed rival offerings from Anthropic, Google, and OpenAI on the CyberGym benchmark, while delivering comparable results at roughly half the cost.
The new model will power Project Perception, a suite of AI agents capable of identifying security vulnerabilities and recommending or implementing code fixes with user approval. The platform is scheduled to enter public preview on August 3 and can integrate with both Microsoft and third-party products.
The announcement marks Microsoft’s first major cybersecurity initiative since Hayete Gallot returned to the company in February as Executive Vice President of Security. Company executives said the technology is designed to help security teams respond more effectively as cybercriminals increasingly use AI to exploit software vulnerabilities.
Microsoft AI CEO Mustafa Suleyman said the company has only tapped a fraction of its cybersecurity data to train the model, suggesting significant room for further improvements. Microsoft CEO Satya Nadella added that combining specialised AI models with security-specific tools and data can improve both performance and cost efficiency.
Microsoft also introduced MDASH, a system that combines MAI-Cyber-1-Flash with GPT-5.4 to identify software flaws that attackers could exploit. The company said the standalone model is not currently available separately and is not subject to US government AI model reviews.
