Sandbox isolation is only slightly important, you don't need to make it fancy, just a plain old VM. The really important thing is how you control capabilities you give for the agent to act on your behalf.
I disagree with this section about WebAssembly:<p>> But the practical limitation is language support. You cannot run arbitrary Python scripts in WASM today without compiling the Python interpreter itself to WASM along with all its C extensions. For sandboxing arbitrary code in arbitrary languages, WASM is not yet viable.<p>There are several versions of the Python interpreter that are compiled to WASM already - Pyodide has one, and WASM is a "Tier 2" supported target for CPython: <a href="https://peps.python.org/pep-0011/#tier-2" rel="nofollow">https://peps.python.org/pep-0011/#tier-2</a> - unofficial builds here: <a href="https://github.com/brettcannon/cpython-wasi-build/releases" rel="nofollow">https://github.com/brettcannon/cpython-wasi-build/releases</a><p>Likewise I've experimented with running various JavaScript interpreters compiled to WASM, the most popular of those is probably QuickJS. Here's one of my many demos: <a href="https://tools.simonwillison.net/quickjs" rel="nofollow">https://tools.simonwillison.net/quickjs</a> (I have one for MicroQuickJS too <a href="https://tools.simonwillison.net/microquickjs" rel="nofollow">https://tools.simonwillison.net/microquickjs</a> )<p>So don't rule out WASM as a target for running non-compiled languages, it can work pretty well!
I also disagree with that.<p>Wasmer can run now Python server-side without any restrictions (including gevent, SQLAlchemy and native modules!) [1] [2]<p>Also, cool things are coming on the JS land running on Wasmer :)<p>[1] <a href="https://wasmer.io/posts/greenlet-support-python-wasm" rel="nofollow">https://wasmer.io/posts/greenlet-support-python-wasm</a><p>[2] <a href="https://wasmer.io/posts/python-on-the-edge-powered-by-webassembly" rel="nofollow">https://wasmer.io/posts/python-on-the-edge-powered-by-webass...</a>
That is a good call out and I missed to consider the options you pointed. When I am back on keyboard I will add an updated note with a link to your comment. Thank you!
OK, let’s survey how everybody is sandboxing their AI coding agents in early 2026.<p>What I’ve seen suggests the most common answers are (a) “containers” and (b) “YOLO!” (maybe adding, “Please play nice, agent.”).<p>One approach that I’m about to try is Sandvault [0] (macOS only), which uses the good old Unix user system together with some added precautions. Basically, give an agent its own unprivileged user account and interact with it via sudo, SSH, and shared directories.<p>0. <a href="https://github.com/webcoyote/sandvault" rel="nofollow">https://github.com/webcoyote/sandvault</a>
Mac Mini + docker for openclaw. Mac Mini is nice because I didnt want to deploy on my local day-to-day machine, otherwise im aware it's not a true security mechanism for an integrated claw.<p>Claude Code local - nothing.<p>Claude Code remote - i just use anthropic's web service. no desire to send my data or use anyone's third party remote sandbox. I would deploy my own before I did that.
Looked into Apples container framework first (for proper isolation) but switched to Docker sandboxes since they switched to mircoVMs too: <a href="https://docs.docker.com/ai/sandboxes/#why-use-docker-sandboxes" rel="nofollow">https://docs.docker.com/ai/sandboxes/#why-use-docker-sandbox...</a>
I'm mainly addressing sandboxing by running stuff in Claude Code for web, at which point it's Anthropic's problem if they have a sandbox leak, not mine.<p>It helps that most of my projects are open source so I don't need to worry about prompt injection code stealing vulnerabilities. That way the worst that can happen would be an attack adding a vulnerability to my code that I don't spot when I review the PR.<p>And turning off outbound networking should protect against code stealing too... but I allow access to everything because I don't need to worry about code stealing and that way Claude can install things and run benchmarks and generally do all sorts of other useful bits and pieces.
I appreciate the details in this, but I also notice it is very machine-focused. When a user wants to sandbox an AI agent, they don’t just want their local .ssh keys protected. They also want to be able to control access to a lot of off-machine resources - e.g. allowing the agent to read github issues and sometimes also make some kinds of changes.
Its worth pointing out another boundary: speculative execution. If sensitive data is in process memory with a WASM VM it can be read even if the VM doesn't expose it. This is also true of multiple WASM VMs running for different parties. For WASM isolation to work the VM needs to be in a seperate process
QubesOS was built to give sandboxes kernel isolation via a hypervisor.<p>It’s not surprising that most people don’t know about it, because QubesOS as a daily driver can be painful. But with some improvements, I think it’s the right way to do it.
[dead]