Microsoft has launched its first artificial intelligence model specifically designed for cybersecurity, MAI-Cyber-1-Flash, alongside a new agentic security system called Project Perception. Announced on July 27, 2026, these tools aim to help organizations defend against increasingly sophisticated AI-powered cyberattacks. Microsoft states that its new offerings achieve superior performance on industry benchmarks and operate at approximately half the cost of existing solutions.
MAI-Cyber-1-Flash is integrated into Microsoft's MDASH harness, a system for identifying and remediating software vulnerabilities. The company claims that this combination achieved a 95.95% score on the CyberGym benchmark, a test designed to measure an AI's ability to reproduce known software vulnerabilities. This score reportedly surpasses leading frontier models such as Anthropic's Mythos, which scored around 84% on the same benchmark, and Google's Gemini. Microsoft also asserts that this configuration is approximately 50% less expensive than its current production setup for MDASH.
Project Perception employs a multi-agent approach, coordinating specialized AI agents to simulate attacks, detect threats, and implement defenses. These agents are categorized into three groups: "red team" agents that probe for weaknesses, "blue team" agents that assess risks, and "green team" agents that remediate vulnerabilities. This system is designed to compress tasks that traditionally take security teams hours into machine-speed operations. Microsoft indicated that while high-impact actions will still require human approval, the system aims to automate much of the detection and remediation process.
Hayete Gallot, executive vice president of Microsoft Security, stated that the cybersecurity landscape has fundamentally changed, with threats evolving from AI-assisted to AI-operated and now autonomous AI attacks. He emphasized that traditional security technologies struggle to match the speed and scale of these advanced threats. Microsoft's strategy with MAI-Cyber-1-Flash involves routing approximately 90% of tasks to this specialized, more cost-effective model, reserving more complex tasks for larger models like OpenAI's GPT-5.4. This approach aims to optimize performance and reduce operational expenses.
The company highlighted its extensive data collection capabilities, processing trillions of security signals daily across various platforms, as a key advantage in developing and training its in-house cybersecurity models. Allie Mellen, principal analyst at Forrester, noted that developing a model based on proprietary data and expertise ensures it is well-suited to reason over Microsoft's specific data and align with its products.
Project Perception will enter public preview on August 3, 2026, and will be integrated initially into Microsoft Defender, with plans for broader rollout across Microsoft's security product portfolio. Microsoft is pricing the system based on consumption, measured in Security Compute Units (SCUs).
This launch occurs amidst a growing concern over the use of AI by malicious actors to automate and scale cyberattacks. Microsoft's announcement also comes as the company faces a class action lawsuit alleging false claims about its Copilot AI assistant's capabilities, including issues with brand positioning, user experience, and competitive performance.