Anthropic hands Claude Code more control but keeps him on the ropes

For AI developers, “vibe coding” currently runs the risk of allowing every action to be child-viewed or model-checked. anthropic he says his latest update to Claude aims to remove that option by letting the AI ​​decide on its own what actions are safe — with some limitations.

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

Automatic mode uses AI guards to review every action before it starts, check for unsolicited risky behavior by the user, and for signs of prompt injection – a type of attack where malicious instructions are hidden in the content the AI ​​processes, causing it to take unexpected actions. Any safe actions will continue automatically, while risky ones will be blocked.

It’s essentially an extension of Code Claude’s existing “dangerous jump permissions” command, which leaves all decision-making to the AI, but adds a layer of security on top.

This feature is based on a wave of autonomous coding tools from companies like GitHub and OpenAI that can perform tasks on behalf of a developer. But it takes it a step further by letting AI decide when to ask the user for permission.

Anthropic didn’t detail the specific criteria the security layer uses to distinguish safe actions from risky ones — something developers will likely want to better understand before implementing the feature widely. (TechCrunch has reached out to the company for more information on this.)

Automatic mode exits behind the launch of Anthropic Claude Code reviewits automatic code reviewer is designed to catch bugs before they reach the codebase and Shipping to Coworkit allows users to send tasks to AI agents to manage work on their behalf.

Automatic mode will roll 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 “isolated environments” — installed in a sandbox that’s kept separate from production systems and limits the potential damage if something goes wrong.