Microsoft Says Its New Security Model Scores 96% on CyberGym. Microsoft Ran the Test.

Microsoft Security announced MAI-Cyber-1-Flash, its first security-specialised model, alongside Project Perception on 27 July, in a post signed by executive vice-president Hayete Gallot. Project Perception is a coordinated red, blue and green agent system and enters public preview on 3 August. Microsoft also announced FORGE Labs, led by vice-president Taesoo Kim, and an external red-team alliance funding 18 university laboratories across six continents.

The claimed result

Microsoft puts its stack at 95.95% on CyberGym — reported as 96% — which it says is 12 points above Mythos. The comparison set it published: GPT-5.5 Cyber at 85.6%, Mythos 5 at 83.8%, GPT-5.6 Sol at 83.6% and Gemini 3.5 Flash Cyber at 83.2%. It also claims "almost 50% of cost savings vs. the current MDASH configuration in market today." Mustafa Suleyman called it "a really quite a remarkable result."

Three problems with the number

First, the benchmark is entirely self-reported: no independent verification, no run count, no error bars. Second, and more substantive, the 96% comes from an orchestrated two-model system — Flash handles roughly 90% of tasks and escalates the rest to GPT-5.4 — benchmarked against competitors' single models. Third, the cost claim is measured against Microsoft's own prior configuration, not against rivals; at least one outlet rendered it as "half the cost of other leading models," which is a materially different and unsupported claim.

What CyberGym measures

Vulnerability reproduction, not exploitation in the wild. A model that reproduces 96% of a benchmark's known vulnerabilities has not solved 96% of security, and Microsoft does not claim it has — but the shorthand travelling in coverage does.

The strategic move underneath

Microsoft has spent two years selling security products built on partner models. It is now shipping its own weights and pricing them against those partners — inside its largest enterprise attach motion. And on the same day, it donated MDASH to Nvidia's new open-model alliance. Building an in-house model, undercutting the vendors it resells, and simultaneously backing the open-weights camp is not three moves. It is one.