No better sandbox: Wasm and the future of AI agent runtimes
summary
Explains why WebAssembly is a stronger sandbox for running AI agents by design, naming concrete protections (no arbitrary memory access, syscall abuse, privilege escalation, kernel escapes, shared-library attacks, ambient filesystem access, or uncontrolled process spawning) and the mechanisms that enforce them (module isolation, capability-scoped imports/exports, virtual filesystem, and Wasm’s Harvard architecture). Uses the Boxer implementation to demonstrate a production-grade, deterministic, inspectable, and hostile-by-default agent runtime. Argues that a Wasm-based runtime is universal across browsers, servers, bare metal, and embedded systems, making safety structural rather than policy-driven.
AI-written summary · about
related talks
48mDon't Trust Anything! Real-world Uses For WebAssembly
35mModern Sandboxing: Using Web Assembly in Java
34mEmbedding WASM in the Future
12mSponsored Keynote: WebAssembly: The future of distributed computing
38mWhy, and how you need to sandbox AI-Generated Code?
34mPlaying Safely in the...
17mA case for Web Assembly
36mSafety-Critical Meets Web-Native: Wasm Revolutionizes Embedded Systems