Anthropic has introduced a new feature known as “auto mode” for its Claude AI system, in a move aimed at reducing the constant oversight developers currently face when using artificial intelligence tools for coding.
The update seeks to address the growing concern around “vibe coding,” where developers must either monitor every action taken by AI systems or risk allowing them to operate without sufficient checks. With this development, the company is attempting to strike a balance by enabling the AI to independently determine which actions are safe to execute within defined limits.
The introduction of auto mode reflects a broader shift within the artificial intelligence industry, where systems are increasingly being designed to act autonomously without waiting for human approval at every stage. While this evolution promises improved efficiency and speed, it also raises concerns about safety and control. Anthropic’s latest feature, currently in research preview, is positioned as an experimental solution that combines automation with built in safeguards to ensure that risky or unintended actions are curtailed before execution.
According to the company, auto mode works by reviewing each action through an internal safety layer that detects potentially harmful behaviour, including prompt injection attacks where hidden instructions could manipulate the AI into performing unintended tasks. Safe actions are executed automatically, while those flagged as risky are blocked, offering a more controlled alternative to earlier tools like the “dangerously skip permissions” command within Claude Code, which granted unrestricted autonomy to the AI.
The feature builds on a growing ecosystem of autonomous coding tools introduced by companies such as GitHub and OpenAI, but goes further by allowing the AI itself to decide when user permission is necessary. Although Anthropic has yet to disclose the exact criteria used by its safety systems, the company confirmed that auto mode will soon be available to enterprise and API users. It will initially support Claude Sonnet 4.6 and Opus 4.6, with developers advised to deploy it within isolated environments to minimise potential risks.