AI 安全與代理
MAI-Cyber-1-Flash Enters Preview, Handling 90% of Vulnerability Analysis with a Small Specialized Model
Microsoft’s MDASH is previewing its first in-house cybersecurity model, MAI-Cyber-1-Flash, handing off only the hardest tasks to GPT-5.4. Multi-model routing reduces scanning costs, but automatically generating exploits and patching production environments also increases risks around permissions and validation.
Microsoft launched the preview of MAI-Cyber-1-Flash on August 3, integrating the specialized cybersecurity model into MDASH, its multi-agent vulnerability analysis system. Derived from the MAI-Thinking-1 family, the model uses a sparse mixture-of-experts (MoE) architecture with 137B total parameters, activating approximately 5B per inference, and supports a 256K-token context window. It is designed not to replace every frontier model, but to handle roughly 90% of code scanning, vulnerability localization, and exploit validation tasks, routing only more difficult cases to GPT-5.4.
Microsoft reports that this combination scored 95.95% on CyberGym while reducing token costs to about half those of the previous best MDASH configuration. The technical significance therefore lies not only in the model’s score, but also in using an inexpensive, low-latency specialized model to process large volumes of candidates, with a more expensive model performing deeper analysis on the small number of highly uncertain cases. MDASH already includes more than 100 agents capable of creating executable proofs of concept, confirming whether vulnerabilities are genuine, and proposing patches.
Project Perception, launched alongside it, divides agents into red, blue, and green teams: the red team identifies attack paths, the blue team assesses risk using asset and identity context, and the green team applies remediation or hardening measures. This creates a closed loop from discovery to response, but also turns faulty patches, software supply chain inputs, and agent permissions into new potential sources of incidents. Microsoft says its test environments are isolated from the internet and that encryption and auditing are provided. Production deployments, however, must connect to real systems, so engineering teams should still require change approvals, rollback-capable deployments, exploit isolation, and independent benchmark reruns. The CyberGym score and reported savings currently come primarily from Microsoft’s own evaluation, and the model does not yet have an independent API, making full external reproduction difficult.