Anthropic defaults Claude Code to auto mode, with safety study showing 89% harm catch rate

Anthropic's shift to auto mode as the default reflects a fundamental rethinking of how code-generation AI should handle safety oversight. The company first tested auto mode in March as a middle ground between speed and control, but the empirical results—showing auto mode catching 89% of harmful actions versus only 13.6% for human review—suggest that automated safety checks outperform human judgment in practice. The discrepancy is striking: users habitually approve 97% of permission prompts when asked for manual approval, a behavior pattern that defeats the intended gate-keeping function. By contrast, auto mode enforces a consistent, rule-based standard for what qualifies as irreversible, destructive, or boundary-crossing, making it more reliable than human intuition for this use case.

The rollout also coincides with new safety layers: prompt injection screening and customizable hard deny rules designed to prevent specific attack vectors like data exfiltration. This layered approach—combining auto-execution with explicit safety guardrails rather than relying on human gatekeeping—marks a shift in how AI companies think about trust and oversight in developer tools.