The anthropic hands of Claude Code are very controlling, but they keep them on the rocks

For developers using AI, “vibe coding” currently comes down to parenting everything or risking the brand not being driven. Anthropic he says his latest update to Claude aims to eliminate that decision by letting the AI ​​decide what’s safe to do automatically – with limits.

The move marks a major shift for the industry as a whole, as AI tools are designed to act without waiting for public approval. The problem is balancing speed and control: too many patrols slow things down, while too few can make behavior dangerous and unpredictable. Anthropic’s new “auto mode,” which is now in research preview — meaning it’s available for testing but not finalized — is its latest attempt to thread the needle.

Auto mode uses AI security to review everything before it starts, looking for unsolicited threats and early injection signals – a type of attack where malicious instructions are hidden in what the AI ​​is planning, causing it to do what it doesn’t want. Any safe actions are automatically performed, while dangerous ones are blocked.

It’s an extension of Claude Code’s existing “skip-allow-allow” rule, which gives all decisions to the AI, but with added security on top.

This feature creates independent scripting tools from companies like GitHub and OpenAI, which can work on behalf of a developer. But it takes a bit of time to shift the concept of when to ask for permission from the user to the AI ​​itself.

Anthropic did not explain in detail how its security system is used to distinguish between safe and dangerous actions – something developers may want to understand before using this feature. (TechCrunch has reached out to the company for further information.)

Auto mode comes from behind the Anthropic setup Claude Code Reviewits independent testers are designed to catch bugs before they hit the codebase, and Cowork Deliverywhich allows users to delegate tasks to AI assistants to perform tasks on their behalf.

Auto mode will be rolled out to Enterprise and API users in the coming days. The company says that it currently only works with Claude Sonnet 4.6 and Opus 4.6, and is promoting the use of a new feature “remote environment” – sandbox settings that are separated from production processes, reducing the damage that can occur if something goes wrong.