Default starting next week

Anthropic announced on Oct. 22 (Oct. 23 Japan time) that "auto mode" will become the default in the web, desktop and mobile apps of its conversational AI "Claude" starting the following week. Claude's official account, @claudeai, made the announcement.

@claudeai · October 23, 2026

Until now, when Claude operated on files or websites, it asked the user for approval for each action. In auto mode, an automated check running in the background judges the safety of each action, letting safe ones through and blocking risky ones. The company is extending a mechanism it had already been using in its programming tool "Claude Code" to its other apps.

Irreversible actions are blocked

Actions that are blocked include those that cannot be undone, those that destroy data and those that reach beyond the work environment. Actions that send confidential information or source code externally are also not allowed through automatically. In Claude Code, if actions are blocked three times in a row, or 20 times in total within the same task, it switches to manual approval. According to the company's research, auto mode blocked 89% of dangerous commands, exceeding the 13.6% caught when humans checked them visually.

It applies to the Pro, Max and Team plans, with new tasks starting in auto mode. For the Enterprise plan for businesses, the API and access through other companies' clouds, it will for the time being remain something users choose to turn on. Configurations that comply with medical information regulations will continue to use manual approval.

Users can switch back to manual

Users can revert by choosing "Manual" in the mode selection. In Team, administrators can set the default mode or disable auto mode. In Enterprise, administrators will be notified by email before it becomes the organization's default.

The automated check does not fully replace the user's judgment. The company recommends switching back to manual approval for actions that move money, send messages in the user's name or involve important files. Just the day before, on the 21st, the company released a research version of "Claude Duet," an always-on agent that acts ahead of users.