As AI agents evolve from simple task executors to autonomous systems capable of discovering and installing third-party tools, the security implications become critical. This session introduces Wassette, Microsoft's open-source WebAssembly-based runtime that provides browser-level security isolation for AI agent tooling through the Model Context Protocol (MCP).

Built on Wasmtime, Wassette uses WebAssembly Components to provide browser-level security with a deny-by-default capability system. Agents can fetch cryptographically signed tools from OCI registries and execute them in isolated sandboxes, requesting permission only for specific resources they need.

This talk demonstrates Wassette in action with GitHub Copilot in VS Code, showing how agents can autonomously install WebAssembly Components for new capabilities like HTTP requests or time services. We'll explore the permission model that allows fine-grained control over network access and system resources, and discuss practical deployment patterns for production cloud-native environments.

Attendees will learn how WebAssembly Components bridge the gap between agent autonomy and security requirements, and see live examples of capability-based sandboxing.