Microsoft Releases MAI-Cyber-1-Flash: 5B Active Parameters and 95.95% on CyberGym

On July 28, 2026, Microsoft AI released MAI-Cyber-1-Flash, its first model built specifically for cyberdefense. Paired with the MDASH scanning harness, it lifted the CyberGym benchmark score to 95.95%, beating the previous industry record.

What is MAI-Cyber-1-Flash?

MAI-Cyber-1-Flash is a sparse Mixture-of-Experts transformer from Microsoft AI with 137B total parameters, of which 5B are active. It has a 256k-token context window and handles text-only input and output. It is a specialized fine-tune of MAI-Code-1-Flash, a lightweight agentic coding model from the MAI-Thinking-1 family that is already built into GitHub Copilot and VS Code. The model does not ship as a standalone API endpoint: it runs only inside MDASH.

• 137B total parameters, 5B active, sparse MoE architecture
• 256k-token context, text-only input and output
• Specialized fine-tune of MAI-Code-1-Flash from the MAI-Thinking-1 family
• 95.95% on CyberGym level 1, versus 88.45% for MDASH in May 2026
• The model handles up to 90% of MDASH tasks; the hardest 10% go to GPT-5.4

Why routing, not the model, sets the record

The headline result comes not from the model itself but from the system around it. MDASH orchestrates more than 100 specialized agents across five stages: Prepare, Scan, Validate, Dedupe, and Prove. CyberGym is a public benchmark of 1507 real vulnerability-reproduction tasks drawn from 188 OSS-Fuzz projects; in the level 1 configuration, the system gets the vulnerable source code and a general description.

MAI-Cyber-1-Flash processes up to 90% of all MDASH tasks and escalates the hardest 10% to GPT-5.4. Microsoft claims this saves 50% compared with the previous stack of GPT-5.4, 5.4 mini, and 5.3 codex. MDASH's previous peak in May 2026, 88.45%, relied only on publicly available models and was already the best public result, about five points ahead of the nearest competitor at 83.1%.

"Replacing 80% of the models inside MDASH moved the harness from 88.4% to 95.95%," says the release write-up from

Microsoft's research team.

Why the model can't write exploits

MAI-Cyber-1-Flash's zeros on the ExploitGym benchmark (0/0/0 across the Kernel, Userspace, and Browser sections) are not a defect but a deliberate Microsoft decision. According to the team, the model was trained for defensive tasks such as patching bugs, not offensive ones such as deploying malicious code.

On real-world work, the pairing has performed convincingly. In May 2026, work involving MDASH produced 16 CVEs in the Windows networking and authentication stack, including four critical remote code execution vulnerabilities. In retrospective testing, the harness recovered 96% of 28 MSRC cases in the clfs.sys driver and 100% of 7 cases in tcpip.sys over a five-year window. MDASH itself was built by the Autonomous Code Security team, which includes members of Team Atlanta, the winner of the DARPA AI Cyber Challenge.

What this means

Microsoft is showing that records in vulnerability discovery come not from a single large model but from smart orchestration of cheap specialized agents: a model with 5B active parameters carries 90% of the work and cuts the bill in half, while a frontier model is called in only for the hardest cases.

Frequently asked questions

What is MDASH?

MDASH is Microsoft's multi-model agentic harness for finding vulnerabilities. It orchestrates more than 100 agents across five stages (Prepare, Scan, Validate, Dedupe, Prove) and scores 95.95% on the CyberGym benchmark.

Why does MAI-Cyber-1-Flash score zeros on ExploitGym?

This is intentional. According to Microsoft, the model was trained only on defensive tasks, such as patching bugs, and was not trained to write exploits or malicious code.

How much cheaper is the new stack than the previous one?

Microsoft claims savings of about 50% compared with the previous configuration of GPT-5.4, 5.4 mini, and 5.3 codex, because MAI-Cyber-1-Flash takes up to 90% of tasks and only the hardest 10% go to GPT-5.4.