The MAI-Cyber-1-Flash cybersecurity model from Microsoft arrived Monday with a benchmark score that should make Google and Anthropic wince, though the fine print is worth reading before anyone declares a winner.
Microsoft unveiled the model at a small event in San Francisco, alongside a new agentic security platform called Perception. The combined launch is the company’s most direct push into specialised AI security to date.
What MAI-Cyber-1-Flash Cybersecurity Model Actually Does
MAI-Cyber-1-Flash is a compact, code-heavy model derived from Microsoft’s MAI-Thinking-1 lineage, built in-house on what the company describes as ‘the highest quality data.’ It is designed to find challenging vulnerabilities in complex codebases, and it powers MDASH, Microsoft’s orchestration harness for software vulnerability identification and remediation.
According to the Official Microsoft Blog, the model handles roughly 90% of all security tasks on its own. The remaining 10%, the genuinely hard cases, get routed to GPT-5.4. That split is also where the cost savings come from: Microsoft claims around 50% lower costs relative to the previous MDASH configuration, which ran on GPT-5.4, GPT-5.4 mini, and GPT-5.3 Codex. The comparison is internal, not a like-for-like against rival vendors.
On the CyberGym benchmark, which Microsoft AI CEO Mustafa Suleyman called ‘the gold standard benchmark for evaluating how systems reason over large codebases to find real vulnerabilities in the code,’ MDASH with MAI-Cyber-1-Flash scored 95.95%. The next-best result, according to daily.dev, belongs to GPT-5.5 Cyber at 85.6%, with Gemini 3.5, GPT-5.6 Sol, and Anthropic’s Mythos 5 clustering around 83% to 84%. That gap is wide enough to be interesting. The benchmark is also one Microsoft uses itself, which is worth keeping in mind.
Suleyman, co-founder of DeepMind, leaned into the results at the event. ‘We’re very very excited to announce our results,’ he said. ‘We have MAI-1 Cyber Flash binded [sic] with GPT 5.4 inside of the MDASH harness, which beats out Gemini, GPT 5.5 Cyber, GPT 5.6 Sol, and Mythos 5 on Cyber Gym, which is the primary benchmark that we all use. The golden benchmark.’ He added: ‘We’re shipping this into production immediately.’
The model draws on more than 100 trillion daily security signals across identity, endpoint, cloud, and network, according to Quartz. MDASH itself coordinates more than 100 specialised agents, each with distinct roles, tools, prompts, and stopping rules. Microsoft’s AI Red Team and an independent third party evaluated MAI-Cyber-1-Flash, and There’s An AI For That reports the model is calibrated for defensive use only, available to verified defenders through MDASH.
Perception and the Crowded AI Security Field
The second launch is Perception, a platform that deploys teams of AI agents across security workflows. Red teams simulate potential attacks, providing context on likely threat actors and the vulnerabilities they might target. Blue teams detect and triage existing bugs. Green teams take, in the words of Microsoft vice president for security Hayete Gallot, ‘corrective actions’ against those bugs.
Gallot framed the whole thing as an arms-race response: a way for enterprise defenders to ‘defend against AI with AI at the scale and speed that the attackers have.’ Dave Weston, Perception’s lead engineer, put the efficiency argument in plainer terms. ‘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. Not only do we discover the issues and prioritize them, but we have detection, posture fixing, and even a code fix.’
Perception will also tap MAI-Cyber-1-Flash for security workflows beyond what MDASH covers, according to the Official Microsoft Blog, making the two launches deliberately intertwined rather than parallel bets.
Neither launch is live for general customers yet. Both will enter preview on 3 November. By then, the field will have had time to respond. Anthropic launched Mythos earlier this year through a limited partner programme called Glasswing, and OpenAI debuted its own security solution in May via a programme called Day Break. Microsoft is arriving with louder benchmark numbers and a tighter product story than either of those early moves, but the proof will be in enterprise adoption figures, not CyberGym scores.
Watch the November preview closely: the first independent red-team results against Perception’s agentic workflows will tell you far more than any self-reported benchmark ever could.
