Overview
- Anthropic will set Claude Code’s classifier‑driven “auto mode” as the default for Pro, Max and Team accounts so the agent proceeds without asking human approval for routine steps.
- Auto mode uses a separate safety classifier on each command that blocks actions judged irreversible, destructive, or aimed outside the user environment and attempts safer alternatives when possible.
- The company reports internal tests with paid users found the classifier flagged a far higher share of harmful actions than human reviewers, and claims teams using auto mode ship more pull requests.
- New safeguards include prompt‑injection screening, customizable hard‑deny rules to stop data exfiltration, and a fallback to manual approvals after repeated blocks in a session.
- Administrators will get controls to pin or disable auto mode and users can switch modes in the app, while Anthropic warns the classifier reduces but does not eliminate risk and recommends human review for high‑stakes production changes.