Microsoft launched its first cybersecurity-specialized model, MAI-Cyber-1-Flash, and a new agentic security platform called Perception at a San Francisco event on Monday, putting the company in direct competition with Anthropic, Google, and OpenAI on AI security. The preview arrives November 3, with Mustafa Suleyman, CEO of Microsoft AI, saying the model ships into production immediately. Microsoft is pitching MAI-Cyber-1-Flash as the top scorer on Cyber Gym, the industry benchmark for AI security tooling.
MAI-Cyber-1-Flash is designed to find vulnerabilities in complex codebases and animates MDASH, Microsoft's harness for software vulnerability identification and remediation. Paired with GPT 5.4 inside MDASH, the model beat Gemini, GPT 5.5 Cyber, GPT 5.6 Sol, and Anthropic's Mythos 5 on Cyber Gym. Microsoft claims the model is both more capable and more cost-effective than those rivals, though specific price and latency figures were not disclosed at the event.
Suleyman, a DeepMind co-founder, framed the results as a step-change moment for the company's in-house security stack.
Key facts
- 01Microsoft launched MAI-Cyber-1-Flash, its first cybersecurity-specialized model, plus an agentic security platform called Perception.
- 02MAI-Cyber-1-Flash paired with GPT 5.4 inside Microsoft's MDASH harness beat Gemini, GPT 5.5 Cyber, GPT 5.6 Sol, and Anthropic's Mythos 5 on the Cyber Gym benchmark.
- 03Both products go into preview on November 3 and will ship into production immediately, according to Mustafa Suleyman.
- 04Perception deploys red, blue, and green agent teams to simulate attacks, triage bugs, and push code fixes.
- 05The launch puts Microsoft directly against Anthropic's Mythos (via Glasswing) and OpenAI's Daybreak, released in May.
Perception is the second half of the announcement and the more ambitious one. The platform deploys agentic red teams that simulate attacks and profile likely threat actors, blue teams that detect and triage existing bugs, and green teams that take corrective actions — including writing code fixes. Perception integrates with MDASH, so vulnerabilities found by MAI-Cyber-1-Flash can be routed straight into remediation workflows.
Hayete Gallot, Microsoft's vice president for security, positioned Perception as a response to attackers who are themselves increasingly using AI.
“defend against AI with AI at the scale and speed that the attackers have.”— Hayete Gallot, Microsoft Vice President for Security
Dave Weston, the lead engineer on Perception, described the practical impact for enterprise security teams. Work that previously required appsec hunters, remediation engineers, and other specialists coordinating across hours now collapses into minutes, he said.
That end-to-end loop — discovery, prioritization, detection, posture fixing, and code fix in one agentic pipeline — is the differentiator Microsoft is selling.
The competitive field is filling in fast. Anthropic launched Mythos earlier this year through a partner program called Glasswing, initially available only to a small group of organizations. OpenAI released its own security solution, Daybreak, in May. Microsoft is the first of the three to pair a specialized security model with a full agentic operations platform under one roof, and the first to publish benchmark wins against the others by name.
The Cyber Gym benchmark claim will be the first thing rivals contest. Benchmarks in security AI are less mature than coding benchmarks like SWE-bench, and Microsoft has not yet published the detailed methodology, per-category scores, or an independent reproduction. The pairing of MAI-Cyber-1-Flash with GPT 5.4 rather than a fully in-house stack also raises questions about how much of the performance is attributable to Microsoft's new model versus the underlying OpenAI model it rides on. Enterprise buyers evaluating Perception during preview will want to see head-to-head results on their own codebases, not just Cyber Gym scores.
For Microsoft, the strategic logic is clear. Security is one of the highest-value verticals in enterprise software, and a fully agentic vulnerability-to-fix pipeline collapses what today is a multi-vendor, multi-role workflow. If Perception's agents can genuinely close bugs in minutes rather than hours, the platform pressures both the traditional application-security tooling market and Anthropic's and OpenAI's newer AI-native competitors. The November 3 preview gives Microsoft a narrow window to lock in reference customers before Mythos exits its Glasswing partner program and Daybreak scales its second half rollout — and it puts every major frontier lab on notice that vertical AI security is now a four-horse race.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
