Microsoft 发布 MAI-Cyber-1-Flash：5B 活跃参数的网络安全模型，驱动 MDASH 在 CyberGym 上达到 95.95%
Microsoft AI Releases MAI-Cyber-1-Flash: A 5B-Active-Parameter Cyber Model That Pushes MDASH to 95.95% on CyberGym
Microsoft 发布 MAI-Cyber-1-Flash，一款 137B 总参数（5B 活跃参数）、256k 上下文窗口的稀疏 MoE 网络安全模型，是 MAI-Code-1-Flash 的微调版本。
微软这个安全模型把漏洞挖掘系统推到95.95%，关键是90%任务由5B模型完成成本砍半，路由设计比模型本身更值得关注，做漏洞挖掘的可以马上跑起来试试。
MAI-Cyber-1-Flash is a transformer with self-attention and sparse Mixture-of-Experts layers. It carries 137B total parameters with 5B active, and a 256k context length. Inputs and outputs are text only.
It is a cybersecurity-specialized fine-tune of MAI-Code-1-Flash, the lightweight agentic coding model already embedded in GitHub Copilot and VS Code. The release describes it as derived from the MAI-Thinking-1 lineage.
Benchmarks
CyberGym is a public suite of 1,507 real-world vulnerability reproduction tasks drawn from 188 OSS-Fuzz projects. Microsoft evaluated at CyberGym’s default level 1 configuration, which supplies vulnerable source and a high-level description.
MDASH running MAI-Cyber-1-Flash alongside GPT-5.4 scores 95.95%. Microsoft frames this as roughly 12 points above Anthropic’s Mythos, and the launch chart places the four competing systems between 83.2% and 85.6%.
When Microsoft first detailed MDASH in May 2026, the harness scored 88.45% on CyberGym using only generally available models. That was already the top public leaderboard score, about five points ahead of the next entry at 83.1%. The research team states the improvement plainly: replacing 80% of the existing models in MDASH moved the harness from 88.4% to 95.95%.
Why the routing is the real product
MDASH manages over 100 specialized agents through five stages: Prepare, Scan, Validate, Dedupe, and Prove. Auditor agents flag findings, debater agents argue exploitability (using disagreement as signal), and the Prove stage executes triggering inputs with ASan for C/C++ targets.
To control frontier model costs at scale, MAI-Cyber-1-Flash handles up to 90% of MDASH tasks, escalating the hardest 10% to GPT-5.4. This routing yields a 50% cost saving over the previous configuration of GPT-5.4, 5.4 mini, and 5.3 codex.
MDASH was developed by Microsoft’s Autonomous Code Security (ACS) team, featuring members from the DARPA AI Cyber Challenge-winning Team Atlanta. In May, MDASH-assisted work generated 16 CVEs (including four Critical remote code execution flaws) in the Windows networking and authentication stack. Retrospectively, it recovered 96% of 28 MSRC cases in clfs.sys and 100% of 7 cases in tcpip.sys over a five-year window.
Performance
The research team present standalone results from a lightweight terminal harness:
| Benchmark | MAI-Cyber-1-Flash |
|---|---|
| CVEBench | 0.314 |
| CyberSecEval4 — Threat Intel | 0.553 |
| CyberSecEval4 — Malware Analysis | 0.33 |
| CRSBench | 0.651 (POV=1200) |
| ExploitGym — Kernel / Userspace / Browser | 0 / 0 / 0 |
The straight zeros on ExploitGym are deliberate, not a defect. Microsoft team states the model was trained to perform defensive tasks such as patching bugs, and not offensive tasks such as deploying malware. A 5B-active model that cannot generate exploits but can drive a 95.95% discovery pipeline is exactly the artifact a defender-only product needs.
How to use it
Key Takeaways
- MAI-Cyber-1-Flash is 137B total / 5B active, a sparse MoE fine-tune of MAI-Code-1-Flash with 256k context.
- 95.95% on CyberGym is a system score — MDASH plus the new model plus GPT-5.4, up from 88.45% in May 2026.
- It handles up to 90% of MDASH tasks, escalating the hard 10% to GPT-5.4 for a claimed 50% cost cut.
- ExploitGym scores are 0/0/0 by design — the model patches bugs, it does not write exploits.
- Access is gated
来源：MarkTechPost（RSS） · marktechpost.com