Background

Developers working with AI‑driven coding tools have faced a trade‑off between speed and control. Existing solutions often require a human to approve each action, or they hand full decision‑making to the model without safeguards. Anthropic’s latest update to Claude Code seeks to bridge that gap by introducing an autonomous mode that can decide which actions are safe to run on its own.

Auto Mode Features

Auto mode is currently available in a research preview, meaning it is ready for testing but not yet a final product. The feature is built on top of Claude Code’s existing “dangerously‑skip‑permissions” command, but adds a safety layer that reviews each potential action. When the AI determines an action is safe, it proceeds automatically; if the action is flagged as risky, it is blocked and the user is notified.

Safety Measures

The safety layer examines actions for two primary concerns: risky behavior that the user did not request and prompt‑injection attacks, where malicious instructions are hidden in the content being processed. By filtering out these threats, Anthropic aims to prevent the model from taking unintended or harmful steps. The company has not disclosed the exact criteria used to differentiate safe from risky actions, but it emphasizes that the feature is intended for use in isolated environments to limit potential damage.

Deployment and Recommendations

Auto mode currently supports Claude Sonnet 4.6 and Opus 4.6. Anthropic recommends running the feature in sandboxed setups that are kept separate from production systems, a practice intended to contain any unintended outcomes. The rollout will be extended to Enterprise and API users in the coming days, expanding access beyond the initial research preview.

Industry Context

This development follows Anthropic’s recent releases of Claude Code Review, an automated code reviewer designed to catch bugs before they enter the codebase, and Dispatch for Cowork, which lets users assign tasks to AI agents. Auto mode represents a broader shift across the AI industry toward tools that can act without waiting for explicit human approval, mirroring similar moves by companies such as GitHub and OpenAI. The challenge remains to balance the speed gains of autonomous execution with the need for robust safety mechanisms.

Looking Ahead

While Anthropic has not provided detailed technical specifications for the safety criteria, developers are likely to seek more transparency before adopting the feature widely. The company’s emphasis on sandboxed environments suggests a cautious approach to deploying autonomous AI in production settings. As the research preview progresses, further refinements to the safety layer and expanded model support may follow, shaping how AI‑assisted coding evolves in enterprise and developer workflows.

Used: News Factory APP - news discovery and automation - ChatGPT for Business