Microsoft launches its first cybersecurity model, plus a new agentic cybersecurity system

Microsoft connected Monday launched its archetypal cybersecurity-specialized exemplary alongside a caller AI cybersecurity level astatine a tiny lawsuit successful San Francisco, taking a large swipe astatine large players successful the abstraction — namely Anthropic, Google and OpenAI.

The institution describes MAI-Cyber-1-Flash arsenic a exemplary that’s built “to find challenging vulnerabilities successful analyzable codebases.” The exemplary is built to animate MDASH, Microsoft’s harness dedicated to bundle vulnerability recognition and remediation.

The caller information level is dubbed Perception, and it’s designed to deploy teams of agents to assistance with and automate assorted information workflows, including identifying and remediating bugs. The level tin besides integrate with MDASH.

The institution claims MAI-Cyber-1-Flash is importantly much almighty (and much cost-effective) than rival models, based connected its show connected an established AI cybersecurity benchmark.

“We’re precise very excited to denote our results,” said Mustafa Suleyman, the co-founder of DeepMind and existent CEO of Microsoft AI. “We person MAI-1 Cyber Flash binded [sic] with GPT 5.4 wrong of the MDASH harness — which beats retired Gemini, GPT 5.5 Cyber, GPT 5.6 Sol, and Mythos 5 connected Cyber Gym, which is the superior benchmark that we each use. The aureate benchmark.”

“We’re shipping this into accumulation immediately,” helium added.

Noting that hackers are progressively utilizing AI successful their cyberattacks, Hayete Gallot, Microsoft’s vice president for security, described Perception arsenic a mode for endeavor defenders to “defend against AI with AI astatine the standard and velocity that the attackers have.”

Perception uses agentic reddish teams, bluish teams, and greenish teams. The reddish teams tin supply elaborate simulations of imaginable attacks — providing discourse astir imaginable menace actors and the apt vulnerabilities that they mightiness exploit. Blue teams are dedicated to detecting and triaging existing bugs, portion greenish teams instrumentality “corrective actions” against those bugs.

Dave Weston, the pb technologist for Perception, described the level arsenic a monolithic ratio upgrade for firm defenders. “We’ve gone from this taking hours and hours of manual enactment from aggregate specialized folks crossed the information enactment — appsec hunters, remediation engineers, you sanction it — and successful minutes, we person a hole for each of this. Not lone bash we observe the issues and prioritize them, but we person detection, posture fixing, and adjacent a codification fix.”

Though AI has offered caller antiaircraft capabilities to companies, its availability to cybercriminals has fixed emergence to a dazzling array of imaginable threats.

Microsoft’s caller information tools, which the institution said volition beryllium disposable successful preview connected November 3, volition participate an progressively crowded tract of AI cybersecurity solutions. Earlier this year, Anthropic launched Mythos, a information level that was released to a tiny coterie of spouse organizations done a programme called Glasswing. OpenAI has also launched its ain information solution successful May done a programme called Day Break.