Anthropic hands Claude Code more control, but keeps it on a leash

Anthropic has unveiled an innovative update to its AI model, Claude, aimed at transforming the way developers interact with artificial intelligence. The newly introduced 'auto mode' allows Claude to autonomously determine which actions are safe to execute, significantly reducing the need for constant oversight by developers. This advancement is part of a larger trend within the tech industry, where AI tools are evolving to operate with greater independence, often bypassing the need for human approval. However, this shift brings the challenge of ensuring a balance between efficiency and control. Too many restrictions can hinder operations, while too few can lead to unpredictable outcomes. Currently in research preview, auto mode employs AI-driven safeguards to evaluate actions before they are carried out. It assesses potential risks and detects prompt injection attacks—where malicious instructions are embedded within the input, potentially leading the AI to execute unintended actions. Safe actions can proceed automatically, whereas risky ones are blocked, enhancing overall security. This feature builds on Claude Code's previous command, which allowed the AI to make decisions independently. However, the new safety layer adds an important level of protection. As autonomous coding tools proliferate, including those from GitHub and OpenAI, this update represents a significant leap by enabling the AI to decide when to seek user permission. While the specifics of how Claude distinguishes between safe and risky actions remain unclear, developers are eager for more details before fully integrating this feature. Auto mode follows the recent launch of Claude Code Review, an automatic code reviewer designed to identify and rectify bugs prior to integration, along with Dispatch for Cowork, which lets users delegate tasks to AI agents. Expected to roll out shortly for Enterprise and API users, auto mode is currently compatible only with Claude Sonnet 4.6 and Opus 4.6. Anthropic advises that the new feature be utilized in 'isolated environments'—sandbox setups that are distinct from production systems, thereby minimizing any potential risks of malfunction.