Microsoft AI launches MAI-Cyber-1-Flash security model
Microsoft AI released MAI-Cyber-1-Flash, a security model tuned for its MDASH harness that it says hits 96% on CyberGym at half the cost of its current best MDASH setup.
Microsoft AI announced MAI-Cyber-1-Flash this week. It is a specialized security model built to run inside MDASH, the company's security harness.
Microsoft says the combined system scores 96% on CyberGym's "any crash" metric. It also says the result beats Mythos on the CyberGym leaderboard and costs 50% of what Microsoft's best current MDASH offering costs. The model runs on H100 GPUs. These are Microsoft's own numbers, and the post gives no breakdown of how they were measured.
The model is meant to handle up to 90% of MDASH tasks. The remaining 10%, which Microsoft calls exceptionally hard problems, go to the largest and most expensive model in its fleet, which in this case is GPT 5.4.
The post also repeats figures from Microsoft's Q4 earnings call. Since last quarter, Microsoft says it has shipped more than a dozen in-house models covering image, voice, transcription, coding and security. It says these models already power widely used Microsoft products and in many cases cut GPU costs by 50–90%. It also says MAI models run on its Maia 200 chip get 40% better performance per watt.
Microsoft frames the work as moving from "tokenmaxxing" to token efficiency. Its argument is that a model tuned to a specific product can match or beat frontier generalist models at much lower cost. It also argues for resilience: the harness, context, memory and action space should be built separately from any one model family, so that each model can be swapped out.
The post does not give pricing, availability details, or the cost baseline behind the 50% claim, and it does not name the products using the new models.