Microsoft launches its first cybersecurity model, plus a new agentic cybersecurity system
TechCrunch Lucas Ropek ● Covered by 4 sources
Microsoft just launched its first cybersecurity AI model plus a new agentic security platform called Perception. It's a direct shot at Anthropic, Google, and OpenAI's own security tools.
Based on reporting by TechCrunch, Lucas Ropek — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Microsoft picked a small event in San Francisco to make a big statement. On Monday, the company unveiled MAI-Cyber-1-Flash, its first cybersecurity-specialized model, built specifically to hunt down tough vulnerabilities buried in complex codebases. It runs inside MDASH, Microsoft's existing harness for finding and fixing software bugs, and it's paired with a new platform called Perception that deploys teams of AI agents to handle security workflows almost entirely on their own.
Mustafa Suleyman, the DeepMind co-founder who now runs Microsoft AI, didn't hold back on the comparisons. He claimed that MAI-1 Cyber Flash, when combined with GPT 5.4 inside MDASH, outperforms Gemini, GPT 5.5 Cyber, GPT 5.6 Sol, and Anthropic's Mythos 5 on Cyber Gym — what he called the industry's go-to benchmark. And Microsoft isn't easing into this. Suleyman said the model is going straight into production.
Perception itself is built around a three-part structure that borrows military-style language: red teams simulate attacks and map out likely threat actors, blue teams detect and sort through existing bugs, and green teams actually fix them. Hayete Gallot, Microsoft's VP for security, framed the whole thing as a necessary response to attackers who are already using AI to move faster than human defenders can. Dave Weston, the engineer leading Perception, put it in blunter terms — work that used to eat up hours of specialized staff time, from app-security hunters to remediation engineers, now gets compressed into minutes, complete with prioritization, detection, and even a code fix.
But Microsoft is hardly alone here. Anthropic already rolled out Mythos earlier this year, though only to a limited group of partners through something called Glasswing. OpenAI made its own move in May with a program named Day Break. So this is less a solo breakthrough than another entry in a fast-filling field, one where every major AI lab now wants to sell both the sword and the shield. Microsoft says its tools will hit preview on November 3, giving enterprise security teams their first real chance to see if the claims hold up outside a stage in San Francisco.
My take — AI-written commentary, not fact-checked reporting
Every major AI lab now sells a cybersecurity product, which says less about brilliant innovation and more about how lucrative the AI-versus-AI arms race has become. Suleyman's benchmark boasting is exactly the kind of chest-thumping that should be treated with skepticism until outside researchers get their hands on these tools. The real test isn't who wins on Cyber Gym — it's whether any of these platforms actually outpace attackers who are just as eager to weaponize the same technology.
Read more about this at: TechCrunch