Microsoft on Monday unveiled Project Perception, an AI cybersecurity system built to defend against AI-driven attacks, aiming to keep pace with both hackers and its technology rivals.

The system, which enters public preview Aug. 3, coordinates three sets of AI agents: red team agents that hunt for paths an attacker could take, blue team agents that determine which risks matter and green team agents that make fixes.

It’s 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 runs in conjunction with OpenAI’s GPT-5.4, which Microsoft reserves for the 10% of tasks it calls exceptionally hard.

Microsoft says the combination scores 96% on CyberGym, a benchmark measuring how well AI systems find real vulnerabilities in large codebases.

The company did not give the model to independent testers before releasing it, according to The New York Times. Microsoft says the model was independently assessed by a third party.

The model is available at launch only to customers of MDASH, Microsoft’s AI-powered tool for finding vulnerabilities in code.

Microsoft CEO Satya Nadella said in a post on X that the initiative is an example of how the company can get better results per dollar by not locking its security systems to a single AI model family.

“This is the benefit of building the harness, context/signals, and action space separate from one model family,” he wrote. “By combining specialized models and data with the right agents, tools, security context, and harness, we can advance the frontier of cost to outcome.”

The initiative was announced Monday morning at an event in San Francisco by Hayete Gallot, the EVP for Microsoft Security, joined by colleagues including Mustafa Suleyman, CEO of Microsoft AI.

In a blog post, Gallot wrote that security needs a new “Cyber Stack,” and that approaches built for a world of human actors cannot keep pace with AI, agents and machine-speed attacks.

In an interview last week for GeekWire’s Microsoft 2.5 series, Gallot said that MDASH was effectively Microsoft’s first step into agentic security.

No system can reason directly over 100 trillion signals a day, so Microsoft is distilling them into a graph that agents can navigate, Gallot said, routing each threat to whichever model handles it best. In practice, this means software can quarantine a device or cut off access on its own.

The announcement comes days after OpenAI disclosed that two of its AI models broke out of a testing sandbox and hacked into Hugging Face, the AI development platform.

Rivals have been more cautious, under government restrictions. 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.

GeekWire is profiling over the next few weeks some of the people and teams that are shaping the evolution of Microsoft in what we’re calling its “Microsoft 2.5” era.

AI has had an impact on just about every tech-product category, but especially security. Attackers are using AI; customers are looking to defend with AI. The goalposts keep shifting. “Agentic security” is now the holy grail, and Hayete Gallot, the newly minted executive vice president of Microsoft Security, is leading the charge toward it.

Gallot, a 16-plus-year Microsoft veteran who rejoined the company in February after a 1.5-year Google detour, replaced Charlie Bell, who came to Microsoft from AWS in 2021 and continues at the company as an individual contributor focused on engineering quality.

“Customers care about two things: solving for security and being able to afford it,” Gallot said when I asked during our interview this week why she came back to Microsoft.

“I am a problem solver. And an engineer at heart (and by training). Security is the most important problem right now — and Microsoft is the only place with all of the puzzle pieces to help our customers.”

Since her return, Gallot hasn’t been shy about shaking things up. As noted recently by The Information, at least nine corporate vice presidents who previously reported to Bell have left the company this year.

“We’re making changes to ensure we’re in the best formation to go after this opportunity,” she acknowledged.

“I’m motivated by doing the right thing for our customers, my teams, and tech outcomes,” she said. “I like to move quickly: days and weeks, not months and years, learning through execution, iterating rapidly, and adjusting based on real customer signals.”

The company isn’t starting from scratch. As of 2021, Microsoft claimed security was a $10 billion business for the company. By 2023, security had reached a $20 billion annual revenue rate, officials said.

Those claims haven’t been without controversy. Microsoft has built a huge business in finding and fixing security problems which some customers felt were of the company’s own making.

Microsoft has a wide-ranging and rather unwieldy security portfolio, encompassing identity management (Entra), endpoint protection (Defender), endpoint management (Intune), security information and event management (Sentinel), and compliance (Purview), among others.

In 2023, Microsoft introduced its Security Copilot set of AI analysis services that integrated with some of its existing security offerings. But a portal-based solution like Security Copilot doesn’t offer the kind of end-to-end coverage that an agentic security platform can, Gallot said.

The problem is that attackers are using agents, too. Customers need real-time insight into what’s happening in their environment, and the ability to act just as quickly, Gallot said.

Agentic security is about “taking the signals and turning them into a graph that is useful,” Gallot said. “If you’re trying to reason about 100 trillion signals, it’s not really effective.” The graph, she said, lets agents pick the right model for each threat and close the loop.

In practice, that means the system can quarantine a device or revoke access on its own, for example, rather than waiting for a human.

[Editor’s Note, July 27: Microsoft announced Project Perception on Monday, the broader agentic security offering anticipated below. Mary Jo Foley wrote further about the announcement for Directions on Microsoft, where she is editor in chief.]

Microsoft’s core existing security products will continue to play a role as the landscape evolves, both spotting the problems and acting on them. Security Copilot isn’t going away in the process: “You’ll have Copilot and you’ll have agentic security,” she said.

The company’s new Agent 365 “control plane” — a central console for tracking every AI agent a company runs — fits in by letting customers see the “blast radius” of an agent, meaning everything a hijacked agent could reach, Gallot said. It’s similar in concept to Zero Trust, the “never trust, always verify” security model that limited how far an attacker could get with a stolen employee login, but applied now to agents rather than people.

So what exactly is this ‘agentic security’ thing? Microsoft has a whole website dedicated to the very topic.

Traditional AI security and agentic AI security are fundamentally different, Microsoft says. Agentic security doesn’t just protect models and training data; it also can protect tools, workflows, memory, connected systems and more. Because agents can take action, the potential positive and negative stakes are higher.

While AI has helped businesses make strides in finding and fixing vulnerabilities, it hasn’t gone much beyond that. Microsoft introduced its multi-model agentic scanning harness (MDASH) as its first step into the agentic security space, Gallot said.

The company used MDASH internally to boost finding and fixing Windows security issues, and it is now making it available to select customers in an expanded preview. MDASH will allow customers to use the best model for the right task to secure all different types of code bases, she said.

Microsoft is rumored to be readying a more comprehensive agentic security offering, of which MDASH is likely just one piece.

Microsoft is far from the only one doing this. AWS, Anthropic, and OpenAI are offering security tools on their platforms, and dedicated security vendors are building their own agentic platforms.

Microsoft has the advantage of scale in the enterprise. The question is whether Gallot and her new leadership team can turn that scale and emerging AI tools into both a bigger business for the company and better protection for its customers.

Rudra “Rudy” Mitra, who spent more than 27 years at Microsoft and most recently led its Purview data-security business, is joining Amazon Web Services as vice president of security services.

Mitra will oversee an AWS portfolio that includes tools such as GuardDuty and Security Hub, which companies use to track security risks across their cloud accounts. AWS recently added AI-specific threat detection to GuardDuty and, perhaps notably given today’s news, extended Security Hub to monitor AI workloads and security inside Microsoft Azure. 

He will report to Chet Kapoor, the former DataStax CEO whom AWS hired last year as vice president of search, security and observability, a role that reports to AWS CEO Matt Garman.

“Rudy brings decades of security experience, a passion for building, and a deep understanding of what customers need as the security landscape continues to evolve,” Kapoor wrote on LinkedIn. 

Mitra joined Microsoft in 1999 straight out of college, working on early efforts to deliver Office as an online service before launching Purview, the company’s data-security and governance product, in 2014. He announced his exit from Microsoft last week, addressing what was next at the time by saying only that there was “more on that soon.”

His departure comes amid a broader reshuffling of Microsoft’s security leadership this year under Hayete Gallot, who returned from Google in February to run the group and has been reshaping its executive ranks in recent weeks and months.

Gallot replaced Charlie Bell, who had joined from AWS in 2021 and continues at Microsoft as an individual contributor focused on engineering quality. She’s been overhauling the group’s product lineup, according to The Information, which reported last week that at least nine corporate vice presidents who reported to Bell have left the company this year.

Rohan Kumar left for Salesforce in June, Vasu Jakkal stepped down after six years. Krishna Kumar Parthasarathy departed this month after 28 years. Joy Chik, president of identity and network access, announced her retirement in April.

On the inbound side at Microsoft, Naseem Tuffaha returned in June to fill the corporate VP role Kumar had left, after nearly two decades at the company and a stint away.

When Gallot arrived, Microsoft named Ales Holecek, a longtime engineering leader, as the security group’s chief architect, reporting to her. David Weston, another veteran Microsoft executive, also reportedly shifted into the security unit earlier this year.