[aniviacat]

Codex has some OS-level sandboxing by default that confines its actions to the current workspace [1].

OpenCode has no sandboxing, as far as I know.

That makes Codex a much better choice for security.

[1] https://developers.openai.com/codex/concepts/sandboxing

[nezhar]

I created VibePod, which allows you to sandbox the agents in containers and monitor their activities. It also supports OpenCode.

https://github.com/VibePod/vibepod-cli