Microsoft adds local file access to Copilot with sandboxed AI agents
Microsoft is expanding Copilot's capabilities by giving the AI assistant access to files on your personal computer, using what the company calls "hybrid intelligence". This system will run some AI processing locally on your device and route complex tasks to cloud-based models such as GPT-6 or Claude. This expansion builds on similar features introduced with Copilot+ PCs in 2024 that have raised privacy concerns among users.
The new functionality uses Microsoft's HydraFusion model router to determine whether tasks should run locally or in the cloud, with local agents operating within sandboxed Execution Containers (MXCs) that have restricted permissions. Microsoft is rolling out these features to GitHub Copilot next week and to the main Copilot application over the coming months. The system will enable advanced search, file organisation and productivity tasks such as finding tax documents and drafting emails, though Microsoft states that Copilot will only access files when explicitly requested by the user.
- Microsoft giving Copilot access to your PC's files through new 'hybrid intelligence' system
- Some AI processing runs locally, complex tasks sent to cloud models
- Rolling out next week starting with GitHub Copilot
New here? Start with this
Copilot is Microsoft's artificial intelligence assistant that can answer questions, write text and help with various tasks. The company is expanding what Copilot can do by allowing it to access files stored on your personal computer, such as documents or emails. To do this safely, Microsoft is using what it calls "hybrid intelligence", which means some of the AI processing happens on your own device rather than everything being sent to company servers.
The system works by having an intelligent router decide whether to process each task locally on your device or send it to more powerful cloud-based AI models. When processing happens locally, the AI runs inside a secured "sandbox" with limited permissions, preventing it from accessing files or systems it shouldn't touch. Microsoft says this design keeps some tasks entirely on your device whilst more complex ones get handled remotely.
This matters because it extends Copilot's usefulness for everyday tasks like finding important documents, organising files or drafting emails. Microsoft is testing these features through its developer-focused GitHub Copilot before rolling them out more widely. The expansion builds on similar local AI features introduced in 2024, which some users have questioned from a privacy perspective.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Advocates argue that local file access with sandboxing represents a necessary and measured step towards making AI assistants genuinely productive for everyday work. By processing sensitive tasks on-device within restricted security containers and intelligently routing only complex operations to cloud services, Microsoft balances capability with privacy protection. Users retain agency through explicit consent, and this hybrid approach pragmatically avoids unnecessary data transmission whilst enabling valuable features like document search and file organisation that modern workers need.
The case against
Sceptics contend that expanding AI access to local files, even with sandboxing, represents a significant security and privacy risk warranting caution. Sandboxes have a documented history of being circumvented by skilled attackers, and the system's complexity expands the attack surface unpredictably. Moreover, 'explicit consent' often obscures what data actually flows to cloud services, and Microsoft's commercial interests in training data create potential misalignment between stated limitations and actual practice—concerns that the earlier Copilot+ PCs rollout, which raised privacy worries, has already illustrated.
Read the full article at the source →
Originally published by The Register as “Microsoft is about to let Copilot loose on your file system”.