Ask HN: The new wave of AI agent sandboxes?
In the last couple of months, several new solutions for sandboxing AI agents have launched (microVMs, WASM runtimes, browser isolation, hardened tool containers, etc.). Curious to hear from people using them in production. Are they working as advertised, or are there still major tradeoffs around security, cost, and performance?
Here's my list of sandboxing solutions launched in the last year alone: E2B, AIO Sandbox, Sandboxer, AgentSphere, Yolobox, Exe.dev, yolo-cage, SkillFS, ERA Jazzberry Computer, Vibekit, Daytona, Modal, Cognitora, YepCode, Run Compute, CLI Fence, Landrun, Sprites, pctx-sandbox, pctx Sandbox, Agent SDK, Lima-devbox, OpenServ, Browser Agent Playground, Flintlock Agent, Quickstart, Bouvet Sandbox, Arrakis, Cellmate (ceLLMate), AgentFence, Tasker, DenoSandbox, Capsule (WASM-based), Volant, Nono, NetFence I would recommend nono - it's practically 0 seconds of latency, developed by the creator of sigstore which is used by google, github, and secures a lot of the open source software supply chain. What I like most is, its like a runtime `nono run ... agent` and there is not managing vms, containers ,mounts. A tool so good that it requires 37 different sandboxes I know there are too many of them, found this benchmarks that my partner did - https://github.com/nickaggarwal/sandbox-test/blob/main/FULL_.... He is planning to add a few more of them. This is very useful. Thanks for sharing! They work, but with tradeoffs. MicroVMs are secure but slower & costly. WASM is fast & cheap but limited. Ultimately, to date, there isn't a perfect solution. A majority of people employ a hybrid solution.