Microsoft’s Security AI Bet and the Cost of Bigger Models
We break down Microsoft’s MAI Cyber One Flash, a 5B-parameter security model that reportedly outperforms larger rivals inside the open-source MDASH harness, and the launch of Project Perception, a multi-agent system for red teaming, blue teaming, and automated remediation.
Then we dig into the economics behind AI infrastructure spending, comparing Microsoft’s specialist approach with Meta’s massive model strategy as earnings from Microsoft and Meta hit after the bell.
Chapter 1
Inside MAI Cyber One Flash and Project Perception
James Turner
Ninety six percent on the CyberGym benchmark. Seriously, let that sink in for a second. We are talking about a five billion parameter model, MAI Cyber One Flash, absolutely running circles around Anthropic's flagship Mythos model. It beat it by twelve full percentage points. And it is doing this while running inside the exact same MDASH harness, that Multi Model Agentic Scanning Harness, that Microsoft open sourced to the Open Secure AI Alliance. It is wild to see how fast we went from them donating the code to now dropping their own custom, in-house silon to run inside it. But here is the thing, it is not just one model. Microsoft is launching Project Perception on August third, which is this massive multi agent security system in public preview. They have got Red Team agents sniffing out attack paths, Blue Team agents doing real time risk assessment, and Green Team agents doing automated clean up. It is like an entire security department in a box. But I am actually a bit torn on this. Microsoft admits in the model card for MAI Cyber One Flash that because of its security first calibration, the safeguards are going to step in more often when a request is ambiguous. I mean, if you are a security engineer trying to analyze a real, nasty exploit payload, and your model keeps saying, oh, sorry, I cannot help you with that, that is unsafe, does that not completely defeat the purpose? It is the classic over refusal trap. You try to make it so safe that it becomes useless for the actual professionals who need to use it to defend their networks.
Chapter 2
The Infrastructure Bill Comes Due
James Turner
And that tension gets even more intense when you look at the economics. Today is Wednesday, July twenty nine, two thousand twenty six. And tonight, after the bell, we have got Microsoft and Meta both dropping their quarterly earnings. This is coming right off the back of Alphabet's Qtwo earnings, where they showed a mind boggling forty four point nine billion dollars in capital expenditure for just one quarter. They are literally paying SpaceX nine hundred twenty million dollars a month just to rent Nvidia GPUs because the supply chain is so choked. Nine hundred twenty million a month! That is insane. So you look at the contrast in strategy here. Meta is out there projected to spend upwards of one hundred forty five billion dollars this year, pushing these massive, generalized open weight models. Meanwhile, Microsoft is dropping this tiny five billion parameter specialist that handles ninety percent of high volume security tasks, cutting their operational compute costs by fifty percent compared to running everything on GPT five point four. As a developer, I have to say, the monolithic model era is feeling a bit tired. The future has got to be these orchestrated constellations of tiny, hyper specific models running in agentic harnesses. It is just so much cheaper and faster. But will Wall Street actually reward this efficiency, or is Meta's brute force, build the biggest brain possible strategy going to win out? I guess we will start to find out tonight when those numbers land. Alright, those earnings are going to be wild. Let us see how the market reacts. Talk to you soon.