AI 安全
Microsoft Routes Vulnerability Analysis Through a Security Model With 5B Active Parameters; MDASH Scores 95.95% on CyberGym
MAI-Cyber-1-Flash is a sparse MoE model with 137B total parameters, but activates only about 5B per token. It handles roughly 90% of MDASH’s workload before passing difficult cases to GPT-5.4. Microsoft says the new configuration cuts costs by nearly half, but the model is currently available only in private preview on Azure AI Foundry, and the benchmark results have yet to be independently reproduced.
Microsoft has unveiled MAI-Cyber-1-Flash, its first internally trained model purpose-built for cybersecurity, and integrated it into MDASH, its agentic vulnerability-analysis system, rather than offering it directly through a general-purpose chat API. The model was derived from a mid-training checkpoint of MAI-Thinking-1, which was first developed into MAI-Code-1-Flash and then fine-tuned for cybersecurity tasks. It uses a sparse mixture-of-experts (MoE) architecture with 137B total parameters, activates about 5B parameters per token, and supports a 256K-token context window.
The deployment strategy is more notable than any single-model ranking. MAI-Cyber-1-Flash handles about 90% of MDASH’s vulnerability discovery and analysis workload, with only the most difficult roughly 10% routed to GPT-5.4. Microsoft reports that this combination scored 95.95% on CyberGym—about 12 points higher than the Mythos-based configuration—while reducing costs by nearly 50% compared with the existing MDASH setup. This suggests that cybersecurity agents may move toward a routing architecture in which a smaller specialist model provides breadth while a frontier model handles the hardest tail-end cases, instead of using an expensive model to scan an entire codebase.
The surrounding Project Perception system further divides the work among red, blue, and green agents: red-team agents identify attack paths, blue-team agents assess risk using environmental context, and green-team agents recommend remediation and hardening measures. This is also where the real engineering risk lies: a benchmark environment can be fully air-gapped, but green-team agents must interact with real codebases and deployment pipelines to deliver value. Teams should watch whether the production release includes change approvals, least-privilege access, reversible patches, complete audit trails, and model-routing logs.
All current figures come from Microsoft, and the CyberGym score combines the model, the MDASH harness, and GPT-5.4; it therefore should not be treated as a measure of MAI-Cyber-1-Flash’s standalone capabilities. The model has neither open weights nor an independent API, so outside researchers currently cannot reproduce its cost, recall, or false-positive rate. After the public preview launches on August 3, cross-project generalization and the safety of automated remediation will be the more important tests.