How do these new tools from Microsoft and NVIDIA address the security risks of running AI agents locally?

The core security feature is Microsoft eXecution Containers (MXC), a sandboxing technology that isolates AI agents and their operations. It creates a policy layer to restrict an agent's access to the broader system, preventing it from interacting with personal files or apps without explicit permission. NVIDIA OpenShell integrates with MXC to provide developers an easy-to-use package for deploying agents securely, helping to mitigate risks like prompt injection attacks.

Build Personal AI Agents on Windows PCs with New Tools from Microsoft and NVIDIA

New Tools for On-Device Agent Development

Microsoft and NVIDIA have announced a collaboration to release a suite of tools for building and securing personal AI agents directly on Windows PCs. Unveiled across COMPUTEX 2026 and Microsoft Build 2026, the initiative aims to meet the demand for on-device AI by providing developers with a standardized stack that includes native security sandboxing, faster inference performance, and direct integration with Windows. This move is notable as it establishes a foundational framework for a new class of applications designed to operate locally and interact with user data safely.

Technical Underpinnings and Performance Gains

The new framework is built around Microsoft's eXecution Containers (MXC), a security primitive that isolates agents to prevent unauthorized system access and mitigate risks like prompt injection. NVIDIA is integrating this technology via its OpenShell runtime, simplifying secure deployment for developers. Key hardware and software advancements support this ecosystem:

Hardware Platform: The new NVIDIA RTX Spark family of desktops and laptops delivers up to 1 petaflop of AI power and 128 GB of memory. Microsoft is also launching a Surface RTX Spark Dev Box preloaded with developer tools.
Inference Speed: Optimizations in inference backends like llama.cpp and vLLM deliver up to 2x performance on certain models through techniques such as Multi-Token Prediction (MTP).
Multi-GPU Support: Both llama.cpp and ComfyUI have been enhanced for multi-GPU RTX PCs, with llama.cpp achieving up to a 1.8x performance increase using tensor parallelism.
Agent Models: Updates to NVIDIA NemoClaw, Hermes Agent, and H Company’s Holo 3.1 models provide improved native Windows integration and performance on NVIDIA GPUs.

Implications for the AI Ecosystem

By delivering a secure, performant, and cohesive toolkit, Microsoft and NVIDIA are lowering the barrier to entry for developing complex agentic workflows that run locally. This initiative directly competes with cloud-centric AI models by empowering development on the installed base of over 100 million NVIDIA RTX PCs. The focus on native security and a streamlined developer experience is positioned to accelerate the creation of personalized AI assistants that are more responsive, private, and deeply embedded in user workflows.

The partnership between Microsoft and NVIDIA signals a strategic push to establish Windows and RTX hardware as the default platform for the emerging on-device agent ecosystem, preemptively building a security framework to manage the inherent risks of autonomous local AI.

>> Verify Original Transmission at NVIDIA