Microsoft's New AI Security Service Can Find Bugs, Simulate Attacks and Write Patches, All in Minutes

Key points

• Microsoft's Project Perception, entering public preview on 3 August, uses multiple AI agents to find software vulnerabilities, simulate attacks, and suggest fixes automatically.
• Microsoft trained its own AI model, MAI-Cyber-1-Flash, specifically to hunt for security flaws in complex software code.
• MAI-Cyber-1-Flash combined with GPT 5.4 scored 96% on the CyberGym security benchmark, 12 points ahead of Anthropic's Mythos 5, at roughly half the cost.
• The system was built by a Microsoft team called FORGE, which includes members of Team Atlanta, the group that won the US government's DARPA AI Cyber Challenge.
• A multi-model design means the system picks whichever AI is cheapest for a given job, rather than routing everything through the most powerful (and most expensive) option.

Most companies that get hacked are not short of security tools. They are short of people. Finding a software flaw, figuring out whether attackers can actually exploit it, writing a detection rule to catch them trying, and patching the underlying code are four separate jobs that might involve four different specialists. Microsoft wants to hand all four jobs to software agents, meaning automated programs that can carry out multi-step tasks on their own.

The new service is called Project Perception. Microsoft announced it this week and plans to open it to enterprise customers, meaning large organisations paying for Microsoft's security products, on 3 August.

How does it actually work?

Perception connects three types of AI agent, each with a specific role. Think of it as a security team running a fire drill on your own systems, automatically, every time a new threat is reported anywhere in the world.

When a security analyst feeds the system a report about a new hacking group, the agents divide up the work. One set scans the organisation's own software and systems to see which parts could be targeted. A second set, playing the role of attacker, probes those parts for weaknesses and ranks them by how dangerous they are. A third set writes detection rules, which are instructions that tell security software what suspicious activity to look for, and drafts patches, meaning software fixes, for the vulnerabilities it found. The system can even submit those patches directly to the internal code repository, the place where a company stores its software source code.

"We've gone from this taking hours and hours of manual work for multiple specialised folks across the security organisation to having a fix for all of this in minutes," said David Weston, a vice president at Microsoft Security who leads the Perception project.

Do you need the most powerful AI to pull this off?

No, and that is the more interesting finding here. Microsoft built its own model, MAI-Cyber-1-Flash, trained from scratch on cybersecurity tasks rather than general knowledge. The model is roughly ten times smaller than GPT 5.4, Microsoft's workhorse AI. Smaller means cheaper to run.

Used inside Microsoft's MDASH harness, a coordinator layer that decides which model handles which task, MAI-Cyber-1-Flash takes on about 90% of routine security jobs. GPT 5.4 handles the remaining tasks that need deeper reasoning. The combination scored 96% on CyberGym, a standard benchmark that tests how well AI systems handle real cybersecurity problems. Anthropic's Mythos 5, a far larger model estimated at 10 trillion parameters (think of a parameter as one tiny dial the AI has learned to tune), scored 12 points lower at a considerably higher cost.

The lesson is one that web security specialists have known for years: a well-designed system beats raw computing power. The right architecture, not the flashiest model, does the heavy lifting.

The team behind the project, called FORGE (Frontier Offensive Research and Generative Exploration), is led by Georgia Tech professor Taesoo Kim. Several members previously competed in the DARPA AI Cyber Challenge, a two-year US government competition for building AI that can defend critical software. Their team, Team Atlanta, won.

What does this mean for ordinary people?

Perception is an enterprise product, so most readers will not interact with it directly. What matters for everyone else is the direction of travel. Security patches and threat detections that once took days may arrive in hours. Organisations that could not previously afford large security teams may get a meaningful level of automated protection.

The flip side, worth watching: the same approach that lets defenders automate threat hunting also lets attackers automate vulnerability discovery. A University of Toronto study published recently showed a self-replicating AI worm built on a free, small model using a similar harness design. The technology is not one-sided.

If you work somewhere that uses Microsoft's security products, ask your IT team when they plan to evaluate Project Perception. The public preview starts 3 August.

Common questions

Does this replace human security staff?

Not entirely. Perception handles the repetitive, time-consuming parts of a security analyst's day, but humans still set priorities, approve major changes, and investigate anything the agents flag as unusual.

Is my personal data affected?

Project Perception is a tool for corporate security teams, not a consumer product. It analyses a company's own systems and software, not personal accounts or individual user data.