Microsoft's MDASH Tops CyberGym Benchmark with 5B Active Params Handling 90% of Tasks

新智元 · wechat · 2026-08-02

Microsoft's multi-agent security system, MDASH, achieved a 95.95% vulnerability reproduction rate on the CyberGym benchmark, significantly outperforming frontier models at half the cost of previous top configurations.

Core Mechanism & Highlights:

Small Model Workhorse: Microsoft's first cybersecurity model, MAI-Cyber-1-Flash (137B total / 5B active params), handles up to 90% of routine queries. Only the hardest 10% are routed to larger models like GPT-5.4, drastically optimizing inference costs.
Multi-Agent Orchestration: The high score isn't from a single model but from a system of 100+ specialized agents collaborating to review code, verify vulnerabilities, debate findings, and write PoCs.
Data Moat: Microsoft leverages over 100 trillion daily security signals and data from 1.6 million customers, creating a barrier to entry competitors cannot easily replicate.

Industry Trend:

Cybersecurity is shifting from "Security Copilots" to "Security Action Systems." Microsoft also announced Project Perception, a larger system using red, blue, and green agent teams to autonomously find, triage, and fix threats, entering public preview on August 3.

Original post →

Microsoft's MDASH Tops CyberGym Benchmark with 5B Active Params Handling 90% of Tasks

More from coding & agent