Claude Code goes autonomous by default, and sandboxing anxiety spikes
Anthropic flipped Claude Code's 'auto mode' to the default, meaning it now takes actions without asking for permission at each step. The HN thread links directly to a sandboxing resource (pleasedonotescape.com) in the second comment, which tells you everything about how people are processing this change. The Docker AI Sandboxes thread, posted the same day, is explicitly pitched as 'disposable, isolated sandboxes for AI agents,' which is a direct response to this exact anxiety.
The pattern: as AI coding agents move from assistants to actors, the security and containment question is becoming urgent. The Claude Code change is not just a UX tweak; it normalizes an autonomous execution model that most developers have not fully thought through.
Docker's sandboxing product is drawing a separate complaint though: it requires a login to use a local dev tool. That friction is being called 'garbage' and 'buzzword reasoning,' which shows that the trust deficit cuts both ways. Developers do not trust the AI to stay in its lane, and they do not trust the vendor to not surveil their local environment.
So what?
If you are shipping an AI agent product, you now need a clear, communicable story about containment. 'We sandbox it' is not enough; you need to explain what the sandbox actually prevents. The login-to-use-local-tooling antipattern is a specific mistake to avoid if you are building dev tools, because the HN audience will roast you for it and they are your early adopters.