Microsoft Unveils Security-Focused AI Model and Platform to Defend Against AI Threats

Amid growing concerns over cyberattacks leveraging artificial intelligence, Microsoft (MS) has unveiled its own proprietary AI models and platforms specialized in security.

On the 27th local time, Microsoft held a security event in San Francisco and introduced "Project Perception," an agent-based security platform.

Project Perception is structured around a collaborative system of three specialized agent teams: a "Red Team" that proactively identifies attack paths, a "Blue Team" that reasons through context to assess risks, and a "Green Team" that implements corrective measures to focus on defense.

The vision is to enable organizations to proactively identify and discover security vulnerabilities before actual attacks take place.

Microsoft explained that the development of this new platform stems from the need for new tools utilizing AI to defend against AI, especially in an era where attacks are executed by autonomous AI.

The platform reportedly works in organic conjunction with MDASH, a software vulnerability management tool previously introduced by Microsoft.

Kim Tae-su, Vice President of Security Research at Microsoft, explained the roles of the two tools, stating, "When MDASH reads code to perform a primary diagnosis of vulnerabilities, Perception understands the customer's environment and context to make decisions and respond regarding how severe a problem a bug actually is."

In particular, Microsoft emphasized that by combining its newly developed security-focused AI model, MAI-Cyber-1 Flash, with OpenAI's GPT-5.4 and applying it to MDASH, it achieved results that outperform top competing models on CyberGym, a security-related benchmark.

The CyberGym score for Microsoft's newly unveiled model is 95.95%, which is more than 10 percentage points higher than OpenAI's GPT-5.5 Cyber at 85.6% or Anthropic's Mythos5 at 83.8%.

Mustafa Suleyman, CEO of Microsoft AI, emphasized, "Through the harness optimized for MDASH and the combined model, we deliver world-class performance at half the cost of competing models."

Suleyman added, "Models themselves will become products within the next two to four years, and there is tremendous room for future improvement because we used only a fraction of the available training data."

This announcement is drawing even more attention as it comes right after an incident where OpenAI's GPT model broke out of its sandbox environment and unauthorizedly hacked the AI platform Hugging Face.

With growing concerns that the threats posed by autonomous AI agents are becoming a reality, Microsoft has essentially presented a "shield," asserting that it can block such attacks.

Project Perception is scheduled to be released to the public as a preview version on August 3.

(Photo: AP, Yonhap News)
※ Please note: This article was translated by AI and may contain errors.