- Microsoft's MAI-Cyber-1-Flash outperforms competitors on the CyberGym benchmark when paired with OpenAI’s GPT-5.4, achieving a score of 95.95%.
- CEO Mustafa Suleyman announced that the model offers world-leading performance at 50% of the cost during an event in San Francisco.
- The public preview of MAI-Cyber-1-Flash will begin next week, with plans to expand Project Perception over time.
- The model will be integrated into Project Perception, which aims to discover and fix vulnerabilities, available in public preview starting August 3.
- Cybersecurity executives believe the new model could help attract more talent to security operating centers (SOCs).
- Combining MAI-Cyber-1-Flash with GPT-5.4 allows the system to achieve a 95.95% score on the CyberGym benchmark, which measures a model's ability to generate working proof-of-concept exploits for known software vulnerabilities.
- Microsoft will make MAI-Cyber-1-Flash available through Azure AI Foundry, utilizing the company's existing customer vetting and GPU provisioning process.
- Weston stated that Microsoft expects organizations to gradually build trust in the technology before allowing agents to operate with greater independence.
- Suleyman mentioned that there is plenty of room for Microsoft to improve the performance of the new model, stating, “We have a unique data set.”
Microsoft's MAI-Cyber-1-Flash, the company's first in-house trained AI model for cybersecurity, has been reported to outperform competitors on the CyberGym benchmark with a score of 95.95% when combined with OpenAI's GPT-5.4. CEO Mustafa Suleyman stated, "We have world-leading performance at 50% of the cost", highlighting significant cost savings for organizations.12910
The model is designed to assist in identifying, prioritizing, and patching software vulnerabilities more efficiently, addressing a critical need in the cybersecurity landscape. It is part of Project Perception, which aims to enhance security operations by utilizing specialized AI agents that mimic human security teams. The public preview of MAI-Cyber-1-Flash is set to begin next week, with plans for further expansion of Project Perception.34
Suleyman emphasized the uniqueness of their data set, stating, "We have a unique data set", which contributes to the model's effectiveness. The introduction of three initial agents—Red, Blue, and Green—will help organizations identify vulnerabilities and deploy patches more rapidly. Weston noted that the model performs roughly 95% of the work done by Microsoft's existing vulnerability-finding system, MDASH.
Despite its promising capabilities, there is still room for improvement, and Microsoft aims to build trust in the technology before allowing greater autonomy for the AI agents. Weston remarked, "We still have to 'earn the right' to make those systems more autonomous".8
“Combining MAI-Cyber-1-Flash with OpenAI's GPT-5.4 achieves a 95.95% score on the CyberGym benchmark, which tests the model's ability to exploit software vulnerabilities. Mustafa Suleyman, CEO of Microsoft AI, emphasized that the model offers world-leading performance at half the cost, aiming to attract more talent to security operating centers.”
CNBC
