Anthropic gives Claude Code more control, but keeps him on a leash

For developers using AI, “vibration coding” right now comes down to taking care of every action or risk letting the model run unchecked. anthropic says Its latest Claude update aims to remove that option by allowing the AI ​​to decide what actions are safe to take on its own, with some limits.

The move reflects a broader industry-wide shift as AI tools are increasingly designed to act without waiting for human approval. The challenge is balancing speed with control: too many barriers slow things down, while too few can make systems risky and unpredictable. Anthropic’s new “auto mode,” which is now in early research (meaning it’s available for testing but not yet a finished product), is its latest attempt to thread that needle.

Automatic mode uses AI protections to review each action before it is executed, checking for risky behavior that was not requested by the user and for signs of rapid injection, a type of attack in which malicious instructions are hidden in the content that the AI ​​is processing, causing it to perform unwanted actions. Any safe actions will be performed automatically, while risky ones will be blocked.

It’s essentially an extension of Claude Code’s existing “permissions dangerously bypass” command, handing all decision-making to the AI, but with a layer of security added on top.

The feature builds on a wave of autonomous coding tools from companies like GitHub and OpenAI, which can execute tasks on behalf of a developer. But it goes a step further by transferring the decision of when to ask the user for permission to the AI ​​itself.

Anthropic hasn’t detailed the specific criteria its security layer uses to distinguish safe actions from risky ones, something developers will likely want to better understand before adopting the feature widely. (TechCrunch has reached out to the company for more information.)

The automatic mode arises after the launch of Anthropic of Claude Code Reviewits automated code reviewer designed to catch bugs before they reach the code base, and Office for Coworkwhich allows users to send tasks to AI agents to handle the work on their behalf.

Automatic mode will be rolled out to enterprise and API users in the coming days. The company says it currently only works with Claude Sonnet 4.6 and Opus 4.6, and recommends using the new feature in “sandboxed environments”: sandbox setups that are kept separate from production systems, limiting potential damage if something goes wrong.