MAI-Cyber-1-Flash, paired with GPT-5.4 inside Microsoft's MDASH harness, outperforms models from Anthropic, Google, and OpenAI on a key benchmark
Bloomberg / Getty Images
Microsoft $MSFT launched MAI-Cyber-1-Flash, its first cybersecurity-specialized AI model, on Monday at an event in San Francisco, alongside a new agentic security platform called Project Perception. The company says the model, when combined with OpenAI's GPT-5.4 inside its MDASH vulnerability management harness, delivers 96% on the CyberGym benchmark — 12 percentage points above Anthropic's Mythos — at 50% of the cost of Microsoft's current MDASH configuration.
MAI-Cyber-1-Flash is intended to shoulder the bulk of routine security work — roughly 90% of tasks — leaving GPT-5.4 to handle only the most demanding cases, the company said. The model is derived from the MAI-Thinking-1 lineage and was built in-house, and Microsoft says it draws on more than 100 trillion daily security signals across identity, endpoint, cloud, and network.
"We have world-leading performance at 50% of the cost," Mustafa Suleyman, CEO of Microsoft AI, said at the event, according to CNBC.
Project Perception, described in a blog post by Hayete Gallot, Microsoft's executive vice president of security, coordinates three classes of specialized agents. Red team agents model how adversaries might move through a system, blue team agents focus on identifying and prioritizing active threats, and green team agents carry out remediation steps to close gaps in defenses. The platform enters public preview on August 3.
Dave Weston, the lead engineer for Perception, said the platform collapses work that previously took hours across multiple specialists. "We've gone from this taking hours and hours of manual work from multiple specialized folks across the security organization — appsec hunters, remediation engineers, you name it — and in minutes, we have a fix for all of this," Weston told TechCrunch.
Gallot, who returned to Microsoft in February to oversee the security division, framed the platform as giving enterprises the means to "defend against AI with AI at the scale and speed that the attackers have," according to TechCrunch.
Microsoft unveiled a family of seven in-house AI models in June, including the MAI-Thinking-1 reasoning model from which MAI-Cyber-1-Flash is derived, as the company has moved to reduce reliance on outside model providers including OpenAI. Microsoft last reported its cybersecurity revenue figures in 2023, putting them above $20 billion annually, and has not updated that figure since.
The new tools enter a market that already includes Anthropic's Mythos, which reached a limited set of partners via a program called Glasswing, and OpenAI's own offering, which debuted in May under a program called Daybreak.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
