Microsoft AI Releases MAI-Cyber-1-Flash: A 5B-Active-Parameter Cyber Model That Pushes MDASH to 95.95% on CyberGym

Microsoft's Autonomous Code Security team just shipped its first model made purely for defense work. It's called MAI-Cyber-1-Flash, and it doesn't stand alone as a product you can just call via an endpoint. Instead it lives inside MDASH, the multi-agent scanning harness Microsoft uses to hunt for vulnerabilities, working alongside GPT-5.4 rather than replacing it entirely.

Under the hood it's a sparse Mixture-of-Experts transformer with 137B total parameters but only 5B active at a time, running a 256k context window and handling text only. It's a specialized fine-tune of MAI-Code-1-Flash, the same lightweight coding model already living in GitHub Copilot and VS Code, and Microsoft traces its lineage back to MAI-Thinking-1.

The headline number is 95.95% on CyberGym, a public benchmark of 1,507 real vulnerability-reproduction tasks pulled from 188 OSS-Fuzz projects, tested at the level where the model gets vulnerable source code plus a high-level description. Microsoft says that beats Anthropic's Mythos by roughly 12 points, with four rival systems clustered between 83.2% and 85.6%. Back in May 2026, when Microsoft first detailed MDASH, the same harness hit 88.45% on CyberGym using only generally available models — already a leaderboard-topping result. Swapping out 80% of the models MDASH relied on is what closed the rest of the gap.

But the score isn't really the point. MDASH is a five-stage pipeline — Prepare, Scan, Validate, Dedupe, Prove — coordinating over 100 specialized agents, with auditor agents flagging issues, debater agents arguing over whether a flaw is actually exploitable, and a Prove stage that runs triggering inputs through ASan for C/C++ targets. MAI-Cyber-1-Flash now handles up to 90% of that workload, kicking only the toughest 10% up to GPT-5.4, which Microsoft says cuts costs by half compared to its earlier setup of GPT-5.4, 5.4 mini, and 5.3 codex.

MDASH already has a track record worth mentioning: in May it helped surface 16 CVEs in Windows networking and authentication code, four of them critical remote-code-execution bugs, and in retrospective testing it caught 96% of 28 known MSRC cases in clfs.sys and every one of 7 cases in tcpip.sys across a five-year span. On a separate terminal-based benchmark suite, the new model posts modest numbers on things like CVEBench and CyberSecEval4, and flat zeros across ExploitGym's kernel, userspace, and browser categories — which Microsoft says is intentional, since the model was trained to patch bugs, not write exploits.

Access to MAI-Cyber-1-Flash is gated for now, which fits the theme: this is a purpose-built defensive tool meant to sit inside a specific pipeline, not a general-purpose model Microsoft wants loose in the wild.