Microsoft is expanding its in-house MAI model family with an AI model specializing in cybersecurity. Just a few days earlier, the company had announced the expansion of its own AI models – now MAI-Cyber-1-Flash follows, a model specifically designed to detect security vulnerabilities in complex codebases. At the same time, the agentic security system Perception is launching, which provides teams of AI agents for various security workflows.
As Microsoft explains in its announcement , MAI-Cyber-1-Flash is a compact, highly code-focused security model. It originates from the MAI-Thinking-1 line and was developed entirely internally on its own training data. According to Microsoft's Model card , the model is based on a Sparse Mixture-of-Experts Transformer architecture with a total of 137 billion parameters, of which only 5 billion are active simultaneously. The context window encompasses 256,000 tokens. Technically, it is a security fine-tune of MAI-Code-1-Flash, which was further developed from a MAI-Thinking-1 Mid-Training checkpoint.
Microsoft is integrating the model into MDASH (Multi-Model Agentic Scanning Harness), a multi-agent system with over 100 specialized agents. These cover the entire workflow from code scanning and validation to patch generation. MAI-Cyber-1-Flash is responsible for around 90 percent of all tasks. The system escalates the most difficult 10 percent to OpenAI's GPT-5.4.
This combination achieves a hit rate of 96 percent on the CyberGym benchmark, according to Microsoft. At the same time, costs are said to be reduced by around 50 percent compared to the MDASH configuration consisting of GPT-5.4, GPT-5.4 mini, and GPT-5.3 codex.
This article was originally published in German . It was translated with technical assistance and editorially reviewed before publication.
The full story
This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.
