For developers working with AI, the current approach to "vibe coding" often involves either closely supervising every step or allowing the model to operate without oversight. Anthropic's latest update to Claude seeks to remove that dilemma by enabling the AI to independently determine which actions are safe to execute—within defined boundaries. This step mirrors a wider industry trend where AI tools are being built to operate autonomously, reducing the need for constant human approval. The central difficulty lies in finding the right equilibrium between speed and control: excessive safeguards can hinder progress, while insufficient ones may lead to hazardous and erratic outcomes.

EMBED_PLACEHOLDER_0

Anthropic's new "auto mode," currently in a research preview—meaning it is open for testing but not yet a finalized offering—represents its most recent effort to address this balance. The mode employs AI-driven safeguards to evaluate each action before execution, scanning for unauthorized risky behavior and indicators of prompt injection. This form of attack involves concealing harmful instructions within content the AI is analyzing, which can trigger unintended actions. Actions deemed safe will proceed automatically, while those identified as risky will be halted.

Essentially, this feature expands upon Claude Code's existing "dangerously-skip-permissions" command, which transfers all decision-making to the AI, but now incorporates an additional layer of safety. It builds on a growing array of autonomous coding tools from firms like GitHub and OpenAI, which can perform tasks directly for developers. However, Anthropic advances this concept by transferring the judgment of when to seek user permission from the person to the AI itself.

The company has not disclosed the exact criteria its safety layer uses to differentiate safe actions from hazardous ones—a detail developers will probably want clarified before embracing the feature broadly.

EMBED_PLACEHOLDER_1

Auto mode follows Anthropic's recent introductions of Claude Code Review, an automated tool for detecting bugs before they enter the codebase, and Dispatch for Cowork, which lets users delegate tasks to AI agents. The auto mode feature will become available to Enterprise and API users in the next few days. Anthropic notes that it currently functions only with Claude Sonnet 4.6 and Opus 4.6, and advises using the new capability in "isolated environments"—sandboxed configurations separated from live systems to minimize potential harm if issues arise.