Microsoft is building an AI security tool that routes tasks across multiple models

Microsoft is developing an AI-powered vulnerability detection tool that works a lot like Anthropic’s Mythos model — but with a critical difference. Instead of relying on a single AI, the system decides which model is best suited for each security task.

Codenamed “Project Perception,” the tool is expected to launch as soon as this month, according to The Information. It represents one of the first major initiatives under Hayete Gallot, who took over as Microsoft’s new security chief in February.

What makes Perception different from existing security tools is its “model router” architecture. Before processing any workload, the system evaluates which AI model — from Microsoft, OpenAI, or Anthropic — is best equipped for the job. It then routes code analysis, vulnerability identification, and automated fix generation to the model that matches each task. The router can also switch between models mid-task if a different capability is needed.

The design doubles as a cost-control strategy. Anthropic’s Mythos has earned a reputation for its advanced hacking capabilities, but deploying it at scale is prohibitively expensive. By distributing requests across multiple models instead of feeding everything through Mythos alone, Microsoft believes it can deliver comparable security capabilities at a much lower price point.

The tool reportedly integrates models from all three AI providers — Microsoft’s own, OpenAI’s, and Anthropic’s — giving it the flexibility to handle everything from proprietary codebase analysis to known vulnerability signature scanning without running up prohibitive costs from any single source.

Gallot has been restructuring Microsoft’s security division since taking over in February, shifting resources toward AI-driven products while cutting back on traditional security offerings. Project Perception is the most visible product to emerge from that realignment so far.