Claude Code's auto mode interception rate is 89%, crushing manual review's 13.6%. Anthropic's test data shows: 1,053 professional testers manually reviewed dangerous commands, with an interception rate of only 13.6%, while Claude Code's auto mode achieved 89%. What's even more critical is that in the prompt injection attack evaluation targeting GPT-5.6 Sol and Claude, Claude's auto mode had a 0% attack success rate, while GPT-5.6 Sol under Codex's auto review mode had a 5.83% success rate. This explains why companies like Adobe and Nuro have already adopted it as the default configuration for production environments—manual review has a "habitual pass-through" issue (97% pass rate), while auto mode is actually safer. Starting August 14, Pro/Max/Team editions will have it enabled by default.