Microsoft is moving deeper into AI-powered cybersecurity with two connected launches: its first purpose-built cyber model and a new platform designed to coordinate teams of autonomous security agents. The new solution, Microsoft MAI-Cyber-1-Flash, is part of this push into advanced AI security.
The model, called MAI-Cyber-1-Flash, was created to find difficult vulnerabilities buried inside large and complicated codebases. It will operate within MDASH, Microsoft’s existing multi-agent system for discovering, validating and remediating software flaws.
Sitting above that is Project Perception, a broader agentic cybersecurity platform that brings together models, threat signals, security context and automated actions.
This is not simply another chatbot for security analysts. Microsoft wants the system to investigate weaknesses, simulate attacks, prioritise risks and help apply fixes before an attacker reaches the same vulnerability.
MAI-Cyber-1-Flash Is Microsoft’s First Dedicated Cyber Model
MAI-Cyber-1-Flash marks Microsoft’s first attempt to build an in-house AI model specifically for cybersecurity work.
The company describes it as a compact, code-heavy model derived from its MAI-Thinking-1 model family. Instead of trying to serve every possible enterprise task, Microsoft trained it around a narrower problem: reasoning through source code and identifying security weaknesses that simpler scanning tools may miss.
That narrow focus matters. Modern codebases can contain millions of lines of code, numerous dependencies and layers of infrastructure that interact in unexpected ways. A vulnerability might not appear inside a single function. It could involve several files, services or memory-management behaviours that only become dangerous when combined.
MAI-Cyber-1-Flash is supposed to handle the bulk of those investigations without requiring Microsoft’s largest and most expensive models for every step. Microsoft says the model can manage as much as 90% of the workload inside MDASH, leaving harder cases to more powerful models.
Microsoft Is Pairing the Model With MDASH
The model does not work alone. Microsoft has integrated it into MDASH, its multi-model vulnerability discovery and remediation harness.
MDASH uses more than 100 specialised agents rather than relying on one AI model and one long prompt. Some agents inspect code. Others challenge the findings, remove duplicates, test whether a suspected flaw can actually be triggered and help generate a repair.
That structure gives Microsoft a way to separate speculation from actionable vulnerabilities.
A basic AI code scanner might flag hundreds of suspicious areas and leave security teams to investigate them manually. MDASH tries to move further by validating whether a flaw is reachable, assessing its possible impact and producing evidence that developers can use.
Microsoft previously reported that MDASH helped its security teams identify 16 vulnerabilities across Windows networking and authentication components. Four carried critical severity ratings and involved remote code execution risks.
The system has since moved into security workflows across Windows, Azure and Microsoft identity products, with findings sent into tools such as Microsoft Defender, GitHub Advanced Security and Azure DevOps.
Microsoft Claims Stronger CyberGym Performance at Lower Cost
Microsoft says the combination of MAI-Cyber-1-Flash and GPT-5.4 inside MDASH achieved a 95.95% result on CyberGym, a benchmark designed to test whether AI systems can reproduce and reason through real software vulnerabilities.
The company rounded that result to 96%, placing the configuration around 12 percentage points ahead of Mythos in its reported evaluation.
Microsoft also claims the new configuration reduces costs by almost 50% compared with the earlier MDASH setup, which used GPT-5.4, GPT-5.4 mini and GPT-5.3 Codex.
Those numbers come from Microsoft’s own testing, so independent validation will still matter. Benchmark performance does not always translate cleanly into enterprise environments, where code is private, incomplete documentation is common and one inaccurate recommendation can create a fresh security problem.
Still, the cost claim may be just as important as the benchmark score. Cybersecurity systems operate continuously. A model that performs well but consumes expensive compute for every alert would be difficult to deploy across thousands of applications and endpoints.
Microsoft’s approach is to reserve expensive reasoning for the small share of cases that genuinely require it.
Project Perception Organises Red, Blue and Green AI Agents
Project Perception expands the idea beyond vulnerability scanning.
Microsoft describes the platform as a continuously learning security system that can perceive risk, reason over context and take action at machine speed while keeping human security teams in control.
The platform organises agents into three broad groups.
Red team agents look for possible attack paths and simulate how an adversary might break into an organisation. Blue team agents investigate signals, assess risk and determine whether suspicious behaviour represents a real threat. Green team agents handle corrective work, such as strengthening configurations, closing exposed routes or helping apply software fixes.
The division sounds almost theatrical, but the purpose is practical. Security work rarely ends when a vulnerability is detected. Teams must understand how it could be exploited, whether an attacker has already used it and what changes will reduce the risk without disrupting normal operations.
Microsoft wants Perception to connect those steps instead of treating them as separate tools and ticket queues.
The Goal Is to Move From Alerts to Action
Security teams already receive more alerts than many organisations can investigate properly. Adding another AI system that simply generates a longer list would not solve much.
Project Perception is built around a different promise: turning security signals into actions.
The platform can combine information from identities, endpoints, applications, cloud infrastructure, company data and AI systems. That shared context should give its agents a clearer picture of how assets relate to each other and which risks deserve immediate attention.
An exposed account on its own may look minor. The same account becomes more serious when it can access sensitive cloud resources, internal code repositories and administrative tools.
Perception is designed to make those connections quickly, then recommend or carry out an appropriate response through Microsoft’s existing security products.
The human analyst still matters. Microsoft says organisations will retain control through governance policies, role-based access, auditing, encryption, tenant isolation and sandboxed environments.
AI Is Speeding Up Both Sides of Cybersecurity
The timing is not accidental.
Cybercriminals can now use AI to search code, modify malware, write convincing phishing messages and automate portions of an attack. That reduces the amount of expertise and time required to probe a target.
Defenders face the same opportunity, though their job is harder in one important way. An attacker only needs to find one opening. The security team must protect every important system without accidentally breaking the business.
Microsoft’s answer is a security stack that operates continuously rather than waiting for scheduled scans and manual investigations. Project Perception is the company’s attempt to make autonomous defense a normal layer of enterprise AI infrastructure and security.
It also pushes Microsoft into more direct competition with AI companies building their own cyber-focused models and research platforms. The market is becoming crowded, and benchmark announcements will not settle the contest. Customers will care about false positives, deployment controls, integration and whether an AI-generated fix can be trusted in production.
Project Perception Is Heading to Public Preview
Microsoft’s official announcement states that Project Perception will enter public preview on August 3, 2026. The company plans to use MAI-Cyber-1-Flash across additional security workflows beyond software vulnerability management.
TechCrunch reported a November 3 preview date, creating a discrepancy between the media report and Microsoft’s own launch post. Microsoft’s August 3 date is the more authoritative timetable at the time of publication.
The larger question begins after the preview opens.
Finding a vulnerability in a benchmark is useful. Finding it inside a live enterprise environment, proving it matters and fixing it without creating new problems is the harder job. Microsoft now believes specialised models and coordinated agent teams can do much of that work in minutes.
That is a serious claim. Project Perception will soon have to prove it outside Microsoft’s own demonstrations.
