Full story
MAI-Cyber-1-Flash Debut
Microsoft unveiled MAI-Cyber-1-Flash, a cybersecurity AI model it says is designed for “using AI to defend against AI,” and it pairs the model with MDASH and OpenAI’s GPT-5.4.
“Should AI models be open or closed”
The company says the combination of MAI-Cyber-1-Flash with MDASH enters public preview on Aug. 3 and is built directly into Microsoft Defender, with a slow rollout to all Microsoft Security products.
Microsoft’s benchmarks posted by the company on Monday say the setup scored 96% on CyberGym, compared with Mythos 5 at 84%.
At a Microsoft briefing on Monday morning, Mustafa Suleyman described the handover process, saying, “MAI-Cyber-1-Flash handles about 90% of the queries.”
Suleyman added that the system then defers about 10% of the queries to GPT-5.4, and he called the CyberGym benchmark result “quite a remarkable result.”
Project Perception Architecture
Microsoft also unveiled Project Perception, an agentic security system that coordinates red team agents, blue team agents, and green team agents to hunt for attacker paths, determine meaningful risk, and apply fixes.
GeekWire reported that the system enters public preview Aug. 3 and that it is based on MAI-Cyber-1-Flash, which Microsoft says does most of the work of larger models at half the cost.
In the company’s own description, Project Perception is built to “turns signals into real-time protections using AI to defend against AI,” and it says it can reason, prioritize and act at machine speed while keeping humans in control.
The Official Microsoft Blog says Perception coordinates three classes of specialized agents, with red team agents identifying potential paths to compromise, blue team agents investigating and determining meaningful risk, and green team agents taking corrective actions.
Microsoft’s blog also frames the approach as a new “Cyber Stack,” saying security needs a system that “continuously perceive[s] risk across the entire digital estate” and takes action at machine speed.
Race, Costs, and Open Debate
Microsoft’s push comes as the AI security race widens, with GeekWire noting that two of the four systems Microsoft benchmarked against, Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol, are limited to small groups of government-approved customers.
“Microsoft has a new AI cybersecurity model called MAI-Cyber-1-Flash, and when it’s combined with agentic security system MDASH and OpenAI’s GPT-5”
In parallel, the debate over open versus closed AI models resurfaced, with BFM citing Jensen Huang’s statement that “Open models strengthen security and cybersecurity, accelerate innovation and diffusion of technologies, and promote sovereignty.”
BFM also reported that Satya Nadella said on X, “We perceive a growing and mistaken sentiment about open models,” and that he added, “It is crucial to understand that open models are essential to safety, to security, to cybersecurity and to innovation.”
The same BFM report described a letter shared by Microsoft CEO Satya Nadella and SpaceX CEO Elon Musk, and it said the signatories argued that open-weight models make state-of-the-art AI “more accessible, more adaptable and widely available.”
BFM further said Scott Bessent warned on X that “sanctions and entity-listing will be considered” if Chinese companies conduct “secret distillation attacks at industrial scale” that cross the red line of IP theft.
