Hey PH! 👋 Maker here. The AI agent space is exciting, but security is still an afterthought in most frameworks. Most AI agents just subprocess.exec() LLM-generated code — that's literally arbitrary code execution. Some use Docker for sandboxing, but that kills performance and portability. So I built BoxAgnts with a WebAssembly sandbox powered by Wasmtime: 🔒 Capability-based security — agents can only invoke explicitly allowed tools ⚡ Near-native execution — no Docker/VM overhead, sub-ms startup 🌍 Cross-platform — compiles to WASM, runs everywhere Also fully open-source (MIT). Try it: grab the binary from Releases. Would you trust an AI agent to execute code on your machine? Why or why not? Would love to hear your thoughts on agent security. https://github.com/guyoung/boxagnts

BoxAgnts

Out-Of-The-Box Secure AI Agent ToolBox in a WASM SandBox

Out-Of-The-Box Secure AI Agent ToolBox in a WASM SandBox