Microsoft touts cost-saving AI model for cybersecurity
Microsoft On Monday, he talked about a new artificial intelligence model to detect cyber security vulnerabilities. The move marks the company’s first major push to revitalize its cybersecurity business since its reinstatement. Google He appointed the unit’s manager, Hayete Gallot.
When paired with OpenAI’s general-purpose GPT-5.4, Microsoft’s MAI-Cyber-1-Flash outperformed Anthropic’s Mythos 5, Google’s 3.5 Flash Cyber and OpenAI’s GPT-5.5 Cyber in the CyberGym benchmark, the company said.
“We have world-leading performance at 50% of the cost,” Microsoft AI CEO Mustafa Suleyman said at an event the company hosted in San Francisco.
The generative model is the software maker’s first model for cybersecurity. He will work on Project Perception, a collection of AI agents to discover and fix vulnerabilities that will be available in public preview starting Aug. 3, according to a statement. blog post From Gallot.
He rejoined Microsoft in February to become vice president of security, a top leader in the category. Amazon cloud manager Charlie Bell was an individual contributor.
Generative AI models have made it easier for attackers to quickly try to exploit newly documented vulnerabilities. Anthropic and OpenAI have released models that can help cybersecurity practitioners defend.
So far in 2026, Microsoft shares are down 19%. “Investor sentiment on Microsoft’s elevated OpenAI exposure has returned to being perceived as a risk, given the consensus view that open-source (Chinese and other) AI models are poised to take share from frontier labs,” Microsoft analysts led by Karl Keirstead wrote in a note to clients Sunday. he wrote. Keirstead recommends buying stocks.
This year, the company announced its own model that can generate code in the GitHub Copilot tool, and has recently been working on a first-party model. Excel spreadsheet program. Microsoft CEO Satya Nadella has a partnership to maintain with OpenAI, but he’s also dedicating his computing power to training models in-house with an eye on spending efficiency.
“By combining proprietary models and data with the right agents, tools, security context, and hardware, we can move from cost to outcome,” Nadella wrote on Monday. x post.
Microsoft hasn’t disclosed the scale of its cybersecurity business since 2023, when it said its annual revenue exceeded 20 billion dollars.
In 2023, Microsoft introduced the Security Copilot assistant for cybersecurity practitioners, featuring OpenAI’s GPT-4. The service now comes bundled with two of Microsoft’s top-tier productivity software suites. Once permission is granted, Project Perception can suggest and implement code changes and connect with non-Microsoft products.
Cybersecurity executives “are looking at this as a way to maybe lower the bar and bring in more talent to staff SOCs and get more people to participate because it’s so limited in the industry right now,” Gallot told CNBC. Companies have security operating centers (SOCs) full of people who monitor threats to information technology systems.
Last week, OpenAI said its models exploited a vulnerability and attacked AI startup Hugging Face’s infrastructure during a test. Hugging Face used a model from Chinese laboratory Z.ai to conduct forensic analysis.
“I think this is a great example of why you should defend with AI against AI bad guys, right?” Gallot said.
There’s plenty of room for Microsoft to improve the performance of the new model.
“We have a unique data set,” Suleiman said in an interview. “We used much less than 1% of this data.”
Microsoft pairs MAI-Cyber-1-Flash and GPT-5.4 for a system called MDASH, which detects bugs that attackers can exploit.
MAI-Cyber-1-Flash “per se is not within the scope of U.S. government reviews and is not currently available as a standalone model,” a spokesperson wrote in an email. A White House official told CNBC earlier this month that he did not approve of the models. Following Anthropic’s release of Mythos 5 in June, the Trump administration implemented and later lifted export controls.
WRISTWATCH: SandboxAQ CEO weighs in on security incidents involving OpenAI and Hugging Face
