Fight AI with AI: Microsoft unveils security-focused model and platform

Microsoft has unveiled a proprietary AI model and platform built specifically for cybersecurity, amid growing concerns about AI-powered cyberattacks.

On Monday (local time), Microsoft held a security event in San Francisco and introduced Project Perception, an agent-based security platform.

Project Perception operates through three specialized agent teams working in concert: a "red team" that proactively identifies attack vectors, a "blue team" that reasons through context to assess risk, and a "green team" focused on taking corrective action and strengthening defenses. The system is designed to detect real-world security vulnerabilities in an organization before an actual attack occurs.

Microsoft said the new platform was developed in response to an era in which attacks are carried out by autonomously operating AI, making it necessary to defend against AI with AI.

The platform works in close integration with MDASH, a software vulnerability management tool Microsoft introduced previously.

Kim Tae-su, Microsoft's vice president of security research, said MDASH reads code and performs an initial diagnosis of vulnerabilities, after which Perception assesses the customer's environment and context to determine how serious a given bug actually is and decides how to respond.

Microsoft said combining its newly developed security-focused AI model MAI-Cyber-1 Flash with OpenAI's GPT-5.4 and applying the combination to MDASH produced results that far outpaced top competing models on CyberGym, a cybersecurity performance benchmark.

The new model scored 95.95 percent on CyberGym, more than 10 percentage points above OpenAI's GPT-5.5 Cyber (85.6 percent) and Anthropic's Mythos 5 (83.8 percent).

Mustafa Suleyman, CEO of Microsoft AI, said the company delivers world-class performance at half the cost of competing models through a harness — a tool for integrating AI models and agents — and a combined model optimized for MDASH.

Suleyman added that within the next two to four years, the model itself will become the product. He said there is also strong room for improvement ahead, as the system has used only a fraction of the available training data.

The announcement drew particular attention coming immediately after an incident in which OpenAI's GPT model broke out of a sandboxed environment and hacked Hugging Face, an AI platform, without authorization. With concerns growing that the threat posed by autonomous AI agents has become a reality, Microsoft is effectively presenting a "shield" — claiming its tools can stop exactly this kind of attack.

Project Perception is scheduled to be released to the public as a preview on Monday.