Microsoft unveils dedicated cybersecurity AI model and agentic platform
The company claims its MAI-Cyber-1-Flash model and Perception platform deliver significant cost savings and performance gains on standard benchmarks, with preview access opening in November.
Microsoft has launched its first cybersecurity-specialised artificial intelligence model, MAI-Cyber-1-Flash, alongside a new agentic security platform named Perception. The announcement was made at a small event in San Francisco on Monday, marking a strategic push into the enterprise security market where the company claims to outperform competitors from Anthropic, Google, and OpenAI.
MAI-Cyber-1-Flash is designed to identify vulnerabilities within complex codebases and serves as the engine for MDASH, Microsoft’s harness for software vulnerability identification and remediation. The model is integrated with larger systems, with Mustafa Suleyman, CEO of Microsoft AI, stating it is bound with GPT 5.4 inside the MDASH harness. Suleyman claimed the combined system beats out Gemini, GPT 5.5 Cyber, GPT 5.6 Sol, and Mythos 5 on the Cyber Gym benchmark, which he described as the primary industry standard.
The accompanying Perception platform utilises agentic red, blue, and green teams to automate security workflows. Red teams simulate potential attacks to provide context on threat actors, blue teams detect and triage bugs, and green teams execute corrective actions. Hayete Gallot, Microsoft’s vice president for security, described the platform as a method to defend against AI-driven threats at the scale and speed that attackers currently possess.
Dave Weston, the lead engineer for Perception, highlighted the efficiency gains for corporate defenders, noting that tasks previously requiring hours of manual work from multiple specialists can now be resolved in minutes. The system provides discovery, prioritisation, detection, posture fixing, and code fixes, aiming to streamline the entire remediation process for security organisations.
The architecture is engineered to handle 90% of security tasks efficiently, reserving larger, more costly models like GPT-5.4 for the remaining 10% of complex issues. This tiered approach results in a 50% cost saving compared to previous MDASH offerings. Microsoft stated the tools would be available in preview on November 3, entering a crowded field that includes Anthropic’s Mythos and OpenAI’s Day Break security solutions.
