Claude Code turns auto mode on by default August 14

Starting August 14, new Claude Code sessions on Pro, Max, and Team plans open in auto mode: every tool call is screened by a separate classifier that blocks irreversible and destructive actions. In Anthropic's tests the classifier stopped 89% of dangerous commands. Manual confirmation caught 13.6%.

The case for it comes from a study with 1,053 paid testers. By the end of long sessions they were catching roughly 5% of those commands, and they approved 97% of all permission requests. Auto mode hands control back to manual confirmation after three blocks in a row, or twenty in a session.

Admins can pin the default or switch auto mode off entirely in managed settings. Anthropic plans to flip the default on Enterprise and in the API within a month.