Microsoft launches Project Perception AI security system

Microsoft launched a new agentic cybersecurity system called Project Perception on July 27, designed to use artificial intelligence to defend against the accelerating threat of AI-powered attacks. The system, which enters public preview on August 3, coordinates teams of specialized AI agents to continuously find, assess, and fix security vulnerabilities.

The announcement, made at a San Francisco event and in a company blog post, introduces Microsoft's first in-house cybersecurity AI model, MAICyber1Flash. This model is integrated with Microsoft's existing MDASH code-scanning system and is central to the new platform. According to the company, on the industry CyberGym benchmark, which covers over 1,500 vulnerability reproduction tasks, the MDASH system running on MAICyber1Flash scored 95.95, outperforming a leading competitor's model. Microsoft executives stated this configuration delivers its performance at roughly half the cost of current leading models by efficiently routing tasks between specialized and frontier AI systems.

Project Perception operates by dividing AI agents into three functional groups. Red team agents proactively probe for weaknesses as an attacker would,...