Microsoft on Monday introduced MAI-Cyber-1-Flash, its first model specifically trained for cybersecurity tasks, and Perception, a new AI platform designed to automate security workflows, at a small event in San Francisco. The company contrasted its offering with competitors including Anthropic, Google and OpenAI.
MAI-Cyber-1-Flash and MDASH
MAI-Cyber-1-Flash is described by Microsoft as a model built to identify challenging vulnerabilities in complex codebases. The model operates within MDASH, Microsoft’s harness for vulnerability identification and remediation in software.
Microsoft said MAI-Cyber-1-Flash is materially more powerful and more cost-effective than competing models based on performance on the Cyber Gym benchmark. Mustafa Suleyman, co-founder of DeepMind and currently CEO of Microsoft AI, said MAI-1 Cyber Flash, bound with GPT 5.4 inside the MDASH harness, beats Gemini, GPT 5.5 Cyber, GPT 5.6 Sol and Mythos 5 on Cyber Gym, which he called the primary and “golden” benchmark.
Suleyman added that Microsoft is shipping the technology into production immediately.
Perception: agent teams for automated security workflows
Perception is designed to deploy teams of agentic red, blue and green teams to assist with and automate security processes, including finding and remediating bugs. The platform can integrate with MDASH.
According to Microsoft, red teams on Perception can generate detailed attack simulations, offering context on potential threat actors and the vulnerabilities they might exploit. Blue teams focus on detection and triage of existing issues, while green teams perform corrective actions against discovered bugs.
Hayete Gallot, Microsoft’s vice president for security, emphasized that attackers are increasingly using AI in cyberattacks and presented Perception as a way for enterprise defenders to “defend against AI with AI at the scale and speed that the attackers have.”
Dave Weston, lead engineer for Perception, described the platform as a major efficiency improvement for corporate defenders. He said tasks that previously required many hours of manual work by multiple specialists—application security hunters, remediation engineers and others—can now be completed in minutes, including discovery, prioritization, detection, posture fixes and even code fixes.
Availability and competitive context
Microsoft said the new security tools will be available in preview on November 3. The announcement places Perception and MAI-Cyber-1-Flash into an increasingly crowded market for AI cybersecurity solutions. Earlier this year Anthropic launched Mythos and made it available to a limited set of partners via its Glasswing program; OpenAI also launched a security initiative called Daybreak in May.
Microsoft’s announcement combines a technical claim—improved vulnerability discovery and automated remediation—with a strategic positioning against other firms developing AI-driven security products. As both attackers and defenders adopt AI capabilities, platforms that can automate detection and response and integrate with vulnerability harnesses like MDASH may play a growing role in enterprise security operations.
