Microsoft is ramping up the AI security race with ‘Project Perception’ and a new in-house model
On Monday Microsoft unveiled Project Perception, an AI cybersecurity program designed to defend against AI-driven attacks, aimed at keeping up with both hackers and its technology rivals.
The system, which goes into public preview on August 3, includes three sets of AI agents: red team agents that hunt down the paths an attacker might take, green team agents that determine which things are dangerous and green team agents that make repairs.
It is based on MAI-Cyber-1-Flash, a new AI model designed specifically for cybersecurity, which the company says does most of the work of larger models at half the cost. It works in conjunction with OpenAI’s GPT-5.4, which Microsoft has reserved for the 10% of tasks it calls the most difficult.
Microsoft says the combination scores 96% in CyberGym, a benchmark that measures how well AI systems detect real vulnerabilities in large code.
The company did not provide the model to independent testers before releasing it, according to the New York Times. Microsoft says the model was independently tested by a third party.
The model is available at launch only for customers of MDASH, Microsoft’s AI-powered code vulnerability detection tool.
Microsoft CEO Satya Nadella said in a post on X that the program is an example of how the company can get better results per dollar by not locking its security systems into a single family of AI models.
“This is an advantage of building a separate harness, core/signals, and action area with one model family,” he wrote. “By combining specialized models and data with the right agents, tools, security contexts, and harnesses, we can drive cost-effectiveness forward.”
The move was announced Monday morning at an event in San Francisco by Hayete Gallot, EVP of Microsoft Security, joined by colleagues including Mustafa Suleyman, CEO of Microsoft AI.
@media (max-width: 600px) { aside.callout { float: none !important; width max:100% !important; margin-left:0 !important; margin-right:0 !important; } aside.callout .callout-img { display:none !important; }}
In a blog post, Gallot wrote that security requires a “Cyber Stack,” and those mechanisms built for the world of human actors cannot keep up with AI, agents and machine speed attacks.
In an interview last week for GeekWire’s Microsoft 2.5 series, Gallot said MDASH was Microsoft’s first step into agent security.
No system can accurately display more than 100 billion signals a day, so Microsoft graphs them for agents to navigate, Gallot said, moving each threat to whichever model handles it best. Basically, this means that the software can isolate the device or cut off its access itself.
The announcement comes days after OpenAI revealed that two of its AI models have moved out of the testing sandbox and into Hugging Face, an AI development platform.
Competitors were more cautious, under government restrictions. Two of the four systems Microsoft compared, Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol, are limited to small groups of government-sanctioned customers.
