Microsoft's Project Perception ships auto-patch security agents Aug 3
The in-house MAI-Cyber-1-Flash model claims 96% on CyberGym at half the cost, and its base coding model is already callable on Azure AI Foundry today.
Copy markdown
A cyber model built to beat the frontier labs
Microsoft's first in-house security model, MAI-Cyber-1-Flash, is a 137B-parameter sparse MoE (5B active, 256K context) tuned for vulnerability work. Inside its MDASH harness Microsoft says it hit 95.95% on CyberGym — about 12 points over Anthropic's Mythos and ahead of OpenAI's GPT-5.5-Cyber.
Half the cost of its old GPT-5.4 stack
MAI-Cyber-1-Flash now handles ~90% of MDASH tasks and routes only the hardest 10% to GPT-5.4, cutting the pipeline's cost roughly 50% versus Microsoft's prior GPT-5.4 + GPT-5.4 mini + GPT-5.3 Codex combo. For teams running large-scale code scanning, cheaper autonomous triage is the real takeaway.
Agents that find, rank and draft the fix
Project Perception coordinates defensive agents that detect vulnerabilities, score severity, and write patches, mirroring a human security team. It's Microsoft's bet on autonomous find-and-fix for software flaws — the direction every code-shipping team should be tracking.
When you can actually touch it
MAI-Cyber-1-Flash is in Azure AI Foundry private preview for approved MDASH customers only, and Project Perception opens to public preview on Aug 3. So it's not a solo-builder tool today — but the public date is close enough to plan around.
The base model you can call right now
If you want a Microsoft model today, MAI-Cyber's base — MAI-Code-1-Flash — is already on Foundry at $0.75/$4.50 per million tokens with a 256K context, and Microsoft cites 51.2% on SWE-bench Pro vs 35.2% for Claude Haiku 4.5. That's the shippable piece of today's news.