Anthropic forces Auto mode in Claude Code after humans approve 86% of bad commands

AI lab Anthropic announced that starting August 14, all new Claude Code sessions on paid plans will default to Auto mode. Instead of asking developers for manual permission before executing every command in the terminal, a background security classifier will automatically evaluate each tool call to block destructive actions.

The decision came after internal research revealed that developers approve 97% of all terminal requests purely out of habit. In a controlled study with 1,053 participants, human coders managed to intercept an embarrassing 13.6% of dangerous commands, while the automated classifier successfully flagged 89% of them.

Vigilance quickly drops off when spamming confirmation prompts. While human participants caught 17% of dangerous actions at the start of a session, their detection rate dropped to just 5% after 50 prompts, effectively turning developers into human rubber stamps.

To make matters worse, power users have been actively sabotaging their own safety systems anyway. Nearly half of active Claude Code CLI users created blanket permission rules for Bash or Python, and 62% enabled options to never ask again for terminal access.

Security benchmarks conducted alongside Apollo Research showed that Auto mode blocked 720 out of 720 indirect prompt injection attacks across Claude Fable 5, Claude Opus 5, and Claude Sonnet 5 without missing a single hit.

Human software engineers have officially outsourced their remaining safety instincts to automated bots, proving that the most vulnerable part of any development pipeline is still sitting right between the keyboard and the chair.

Source: Anthropic