Microsoft AI launches its first cybersecurity model

Microsoft AI has announced MAI-Cyber-1-Flash, its first in-house cybersecurity model. The model is built into MDASH, Microsoft's multi-agent harness for finding and fixing software vulnerabilities.

The company describes MAI-Cyber-1-Flash as a compact, code-heavy model from its MAI-Thinking-1 line. It is meant to handle up to 90% of tasks in MDASH. The remaining 10%, which Microsoft calls exceptionally hard, go to GPT-5.4.

Microsoft says the combined system scores 96% on CyberGym. That figure uses the benchmark's "any crash" score, and Microsoft says it beats Mythos on the CyberGym leaderboard. The company also claims the system outperforms Gemini and GPT. The 96% belongs to the full MDASH system, not to MAI-Cyber-1-Flash on its own. The post also carries a "clarification on CyberGym scores," updated August 13, 2026, but the provided text does not say what was clarified. All benchmark results come from Microsoft.

On cost, Microsoft says the new setup is 50% cheaper than its current best MDASH configuration, which uses GPT-5.4, GPT-5.4 mini and GPT-5.3 Codex. More details are in a technical report.

Microsoft attributes the results partly to its security data. It cites more than 100 trillion security signals a day and operational data from 1.6 million customers, and calls security work a live reinforcement learning loop. MDASH runs more than 100 agents built by Microsoft's security experts.

Microsoft also launched Perception, an agentic security system for monitoring and patching, which it says will soon use MAI-Cyber-1-Flash. The company says the model was tested by its AI Red Team and assessed by an unnamed third party.

Microsoft did not say whether the model will be available outside MDASH.