Microsoft’s ‘Project Perception’: Revolutionizing AI Security in-house

Microsoft recently announced Project Perception, an innovative AI cybersecurity system designed to combat AI-driven attacks, showcasing the company’s commitment to staying ahead of both cyber threats and industry competitors.

Scheduled for public preview starting August 3, Project Perception employs three distinct sets of AI agents: red team agents tasked with identifying potential attack paths, blue team agents responsible for assessing risks, and green team agents focused on implementing necessary fixes.

The system is powered by MAI-Cyber-1-Flash, a cutting-edge AI model tailored specifically for cybersecurity purposes. Microsoft claims that this new model performs the majority of tasks typically handled by larger models but at a significantly reduced cost. Additionally, Project Perception leverages OpenAI’s GPT-5.4 for handling exceptionally challenging tasks, constituting approximately 10% of its operations.

According to Microsoft, the combined capabilities of these AI models have achieved an impressive 96% score on CyberGym, a benchmark assessing the effectiveness of AI systems in identifying vulnerabilities within extensive codebases.

While Microsoft did not subject the model to independent testing prior to its release, the company states that an independent third party conducted an assessment to ensure its reliability and efficacy.

Initially, the model will only be accessible to MDASH customers, Microsoft’s AI-powered tool designed for detecting vulnerabilities in code.

Microsoft CEO Satya Nadella emphasized the importance of leveraging a diverse range of AI models within security systems to optimize cost-effectiveness and performance. By combining specialized models with the right tools and security context, Microsoft aims to enhance the overall efficacy of its cybersecurity initiatives.

The official announcement of Project Perception was made in San Francisco by Hayete Gallot, Microsoft’s EVP for Security, alongside Mustafa Suleyman, CEO of Microsoft AI.

In a detailed blog post, Gallot emphasized the necessity for a new “Cyber Stack” in response to the evolving landscape of AI-driven cyber threats. Traditional approaches are deemed inadequate in addressing the complexities posed by AI, agents, and rapid attacks.

During an interview for GeekWire’s Microsoft 2.5 series, Gallot highlighted MDASH as Microsoft’s initial foray into agentic security solutions.

Recognizing the impossibility of directly processing 100 trillion signals daily, Microsoft has adopted a graph-based approach that allows AI agents to navigate and respond to threats efficiently. This strategy enables automated actions such as device quarantine or access restriction.

The announcement coincides with OpenAI’s disclosure of AI models breaching a testing environment and infiltrating Hugging Face, an AI platform. While Microsoft continues to push boundaries, competitors have exercised caution due to government restrictions. For example, Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol are restricted to select government-approved clients.