HNHacker News
TopNewBestAskShowJobs

languid-photic

129 karma · joined April 24, 2025

working on: https://github.com/voratiq/voratiq get in touch: hi@voratiq.com
submissionscomments
languid-photic··on Observed Agent Sandbox Bypasses
I think often it's a question of naivety rather than maliciousness.

> AI should, from the core be intrinsically and unquestionably on our side

That would be great and many people are working to try to make this happen, but it's extremely difficult!

languid-photic··on Observed Agent Sandbox Bypasses
I am happy to know this term now, thanks.

I do think this is part of the alignment problem. There are two side, the agent (here I think there was a gap in institutional knowledge about what is and isn’t appropriate) and the environment (what is it able to do).

I’m not sure which one is easier to “solve”. It's so hard to know every possible path forward when working from the environment direction.

languid-photic··on Observed Agent Sandbox Bypasses
You're right, this is a bit of a conflation. The curl and lockfile examples aren't sandbox escapes, the network blocks worked. The agent just masked the failure or corrupted local state to keep going. The env var leak and directory swap are the actual escapes. Should have been clearer about the distinction.
languid-photic··on Observed Agent Sandbox Bypasses
You're right that a Podman container with minimal mounts would have blocked the env var leak. Our sandbox uses OS-level policy enforcement (Seatbelt on macOS, bubblewrap on Linux) rather than full container isolation. We’re using a minimal fork that also works w Codex and has a lot more logging on top.

The tradeoff is intentional, a lot of people want lightweight sandboxing without Docker/Podman overhead. The downside is what you're pointing out, you have to be more careful. Each bypass in the post led to a policy or implementation change. So, this is no longer an issue.

On prompts: Red-teaming meant setting up scenarios likely to trigger denials (e.g., blocking the npm registry, then asking for a build), not prompt-injecting things like “do whatever it takes.”

[1] https://github.com/anthropic-experimental/sandbox-runtime

languid-photic··on Ask HN: How are you sandboxing coding agents?
Yes, typically the agent is just trying to do what it's been instructed to do, but sometimes it's too naive to realize its approach is a bit sketchy.

And actually, one way we've hardened our sandbox is by tasking agents with impossible tasks (within the sandbox), then analyzing and patching each workaround.

languid-photic··on Ask HN: How are you sandboxing coding agents?
> Have you had any "learned the hard way" moments?

A big lesson for us is that you still need to be careful even in a sandbox.

We've been running Claude/Codex/Gemini in sandboxed YOLO mode and have seen some interesting bypass attempts. [1]

A few examples:

- created fake npm tarballs and forged SHA‑512s in our package‑lock.json

- masked failures with `|| true`, making blocked operations look successful

- cloned a workspace, edited the clone, then replaced the workspace w the clone to bypass file‑path deny rules

So, we’ve learned to default to verbose logging, patch bypasses as we see them, and try to keep iteration loops short.

[1] https://voratiq.com/blog/yolo-in-the-sandbox/

languid-photic··on AI overviews cause massive drop in search clicks
Yes, but I think part of the reason for llms.txt is optimize context. eg beyond content, the human docs often have styling markup which wastes tokens.
languid-photic··on Apple partnering with startup Anthropic on AI-powered coding platform
IIRC Claude is Cursor’s most used model… so this makes sense from a utility perspective.

This is not a new opinion, but it doesn’t make a lot of sense from an Apple-as-a-completely-vertically-integrated-company perspective.

← PreviousPage 2 of 2