Microsoft’s First Cyber Model MAI-Cyber-1-Flash Takes Aim at Anthropic, Google, and OpenAI
Quick Read
- Microsoft has launched MAI-Cyber-1-Flash, its first in-house cybersecurity model
- A new agentic platform called Project Perception rolls out three types of AI agents: red, blue, and green teams
- Microsoft claims the model beats rivals from Anthropic, Google, and OpenAI on the CyberGym benchmark
- Perception enters public preview on August 3, starting inside Microsoft Defender
- The hardest 10% of tasks still get handed off to OpenAI’s GPT-5.4
Microsoft just made its biggest move yet in AI security, launching its first cyber model and taking direct aim at Anthropic, Google, and OpenAI, according to TechCrunch, which reported the launch happened at a small event in San Francisco.
The model, MAI-Cyber-1-Flash, built to hunt down hard-to-find vulnerabilities in messy codebases. It plugs into MDASH, Microsoft’s existing harness for identifying and fixing software vulnerabilities.
Alongside it, Microsoft launched Project Perception, an agentic platform that splits work across three agent types: red teams that simulate attacks, blue teams that detect and triage threats, and green teams that patch things up, based on a report describing all three teams running in parallel inside the same harness. It’s essentially a full security team, automated and running around the clock.
On Performance, Microsoft Isn’t Holding Back
The company says MAI-Cyber-1-Flash scored 96% on the CyberGym benchmark, 12 points ahead of Anthropic’s Mythos. Still, the model handles roughly 90% of tasks alone, while the hardest 10% get routed to OpenAI’s GPT-5.4, a slightly awkward reminder of how demanding top-tier security work really is.
Microsoft AI CEO Mustafa Suleyman called this just the beginning, pointing to Microsoft’s data, harness, and expertise as giving it an edge in training faster, cheaper models. Worth noting: Microsoft isn’t first here. Anthropic previewed Mythos in April, and Google launched Gemini 3.5 Flash Cyber just last week.
Perception enters public preview August 3 inside Microsoft Defender, with MAI-Cyber-1-Flash landing on Azure AI Foundry the same day.
