Microsoft updates GitHub Copilot with local inference and VS Code sandboxing

Microsoft has announced significant updates to GitHub Copilot, introducing support for local AI model inference. By the end of the month, developers will be able to switch between cloud-based and device-side models, either automatically or manually, within tools such as Visual Studio Code and the GitHub Copilot CLI.

To facilitate this, Microsoft introduced the MAI Code 1.1 Flash “mixture-of-experts” model. Optimized with quantization and speculative decoding, the model demonstrated high performance on the Surface Laptop Ultra, achieving decoding throughput between 40 and 63 tokens per second.

In conjunction with these capabilities, Visual Studio Code version 1.141 introduces sandboxing for Copilot. This security feature provides an isolation layer that limits an AI agent's access to specific files and network resources. The update is designed to mitigate risks such as prompt injection attacks, unreliable dependencies, and errors during the execution of tasks in sensitive projects.