AI 安全與代理系統
Microsoft Uses a Small Cybersecurity Model for 90% of Scanning, With MDASH Escalating Only the Hardest Tasks to GPT-5.4
Microsoft has integrated its first in-house cybersecurity model, MAI-Cyber-1-Flash, into the multi-agent vulnerability-scanning system MDASH, allowing a cheaper specialized model to handle most of the workload. The company says the hybrid configuration achieved 95.95% on CyberGym, but the benchmark measures vulnerability reproduction and does not directly demonstrate the safety of automated patching in real-world environments.
Microsoft unveiled MAI-Cyber-1-Flash on July 27. Derived from MAI-Thinking-1, it is a compact model optimized for code and cybersecurity data. Rather than independently replacing frontier models, it is embedded in MDASH’s multi-model agent framework: the specialized model handles about 90% of vulnerability analysis, while only the most difficult 10% of tasks are routed to GPT-5.4. This tiered inference approach applies model routing directly to cybersecurity economics. Microsoft says it can reduce costs by 50% compared with the previous configuration of GPT-5.4, GPT-5.4 mini, and GPT-5.3 Codex.
MDASH itself comprises more than 100 agents that perform tasks including code comprehension, vulnerability hypothesis generation, exploit validation, and patching. The new hybrid system scored 95.95% on CyberGym, which contains 1,507 known-vulnerability reproduction tasks—about 10 to 13 percentage points higher than the Mythos, Gemini, and GPT comparison configurations listed by Microsoft. The company also introduced Project Perception, which initially assigns three agent teams—Red, Blue, and Green—to identify weaknesses, assess risk, and generate patches, respectively. Microsoft plans to expand MAI-Cyber-1-Flash into additional security workflows.
Deployers should note that the 95.95% figure represents the performance of the complete MDASH system together with GPT-5.4 and cannot be attributed to a single model. The prompts, tool budgets, and inference costs used in the comparison configurations have also not been fully disclosed. Microsoft will not release the model weights; instead, access will be provided through Azure AI Foundry, subject to customer review and compute allocation, with a public preview scheduled to begin one week after the announcement. Microsoft says the platform provides tenant isolation, role-based access controls, auditing, and network-isolated sandboxes. Engineering teams must nevertheless validate false positives, patch regression testing, permission boundaries, and human-approval workflows for cases in which agents generate executable exploit code.