Microsoft introduces its first cybersecurity AI model at half the cost of competing models
Microsoft has introduced MAI-Cyber-1-Flash, its first AI model built specifically for cybersecurity. Designed to identify complex vulnerabilities in large codebases, the model works with GPT 5.4 inside Microsoft’s MDASH system to support vulnerability discovery and remediation. Microsoft claims this setup outperformed Gemini, GPT 5.5 Cyber, GPT 5.6 Sol, and Mythos 5 on the Cyber Gym benchmark, while operating at roughly half the cost.
The company also announced Perception, an agentic cybersecurity platform that coordinates red, blue, and green AI teams. Red teams simulate attacks and analyze potential threats, blue teams detect and triage vulnerabilities, and green teams handle remediation, security configuration changes, and code fixes. Microsoft says the platform can reduce work that previously required several specialists and hours of manual effort to minutes.
Perception integrates with MDASH and uses MAI-Cyber-1-Flash within its agent-based system. Microsoft is positioning the tools as a way for organizations to respond to AI-assisted cyberattacks at greater scale and speed. Perception will enter preview on August 3.
- AI Chatbot
- Freemium
- Proprietary
Other apps mentioned in this article
Related news
External links
- Introducing MAI-Cyber-1-Flash inside MDASHMicrosoft AI • Official source
- Microsoft Says Its New Cybersecurity AI Beats Industry Leaders at Half the CostCNET
- Microsoft launches its first cybersecurity model, plus a new agentic cybersecurity systemTechCrunch
- Microsoft launches AI cybersecurity model, agentic defense platform to cut enterprise security costsVentureBeat
Comments
Citing Microsoft, it only performs better on CyberGym (which is not a "gold standard benchmark" as Microsoft is pretending), and is 50% cheaper than its current weird stack (composed of different ChatGPT versions melted altogether, since Microsoft still got IP rights on ChatGPT models), and is the new addition of the great MAI (reading "Microsoft AI") models family that nobody cares about (that ones that are deeply integrated into all possible Microsoft services).
The preview will be on August 3, not November (since OpenAI and Anthropic will certainly release few more models since then, or fail their IPO and choke by their operational debts and lack of funding, whatever).
I’ve fixed the date in the Perception preview. Thanks for the heads-up!