Microsoft Corporation has made a significant breakthrough in cybersecurity by releasing its first specialized language model — MAI-Cyber-1-Flash. This model has already been integrated into the MDASH vulnerability search system, which, according to the developers, demonstrates impressive results, surpassing well-known solutions such as Mythos 5 and GPT-5.5-Cyber.
Key Metrics and Comparison with Competitors
According to my analysis, the combination of MAI-Cyber-1-Flash and MDASH scored 95.95% on the CyberGym benchmark. This test challenges AI agents to reproduce 1,507 known vulnerabilities across 188 open-source projects. For comparison, GPT-5.5 Cyber achieved 85.6%, Mythos 5 — 83.8%, and GPT-5.6 Sol — 83.6%. Notably, MAI-Cyber-1-Flash is 50% cheaper than the existing MDASH configuration based on GPT-5.4, 5.4-Mini, and 5.3-Codex.
Architecture and Hybrid Approach
MAI-Cyber-1-Flash is built on a hybrid approach: the model independently handles up to 90% of tasks, while the remaining 10% of the most complex queries are redirected to GPT-5.4. This allows for significant token savings and increased overall efficiency. The MDASH ecosystem involves over 100 specialized AI agents that handle code auditing, verification of findings, and creation of Proof-of-Concept — working examples that confirm the presence of vulnerabilities.
Availability and Prospects
In the initial phase, MDASH is available to a limited number of users through the Microsoft Security Exposure Management service on the Defender portal. Clients can scan repositories on GitHub, analyze discovered vulnerabilities ranked from "unlikely" to "confirmed," and use Defender CLI to automatically generate code fix suggestions.
My expert opinion: This move by Microsoft marks a new era in cybersecurity, where specialized AI models not only surpass general-purpose counterparts in accuracy but also significantly reduce operational costs. However, it is worth noting that the benchmark data has not yet been presented in a public ranking, leaving room for independent verification. Nevertheless, the hybrid approach and cost-effectiveness of MAI-Cyber-1-Flash could become the industry standard.