Tools
AI Minute Newsroom
2026-08-14
From today Claude Code stops asking permission — a second model decides what is too dangerous to run
Starting 14 August, auto mode is the default permission setting for new Claude Code sessions on Pro, Max and Team plans. Instead of a prompt before each shell command, a separate classifier screens every action and only interrupts for things it judges irreversible, destructive, or aimed outside your own environment. In a controlled study with 1,053 paid testers, the classifier blocked 89% of dangerous commands while people clicking through the old approval prompts caught 13.6%. If it blocks three commands in a row, or twenty across one session, Claude Code falls back to manual approval. Enterprise, API and cloud-platform users stay opt-in for about another month. Shift+Tab still switches modes, admins can pin a default or turn auto mode off for a team, and anyone who had already chosen a default gets a one-time prompt.
Why it mattersThe old design treated a human reading each prompt as the safety layer. Anthropic's own numbers say that layer barely worked: users approved 97% of permission prompts, and their catch rate fell the longer a session ran — roughly 17% early on, around 5% after fifty prompts. Handing the check to a classifier that does not get bored is a measurable improvement, and it is the first time a major coding agent has published a study saying the human in the loop was the weak part. It also moves the last human veto a step further from the keyboard. If you run Claude Code on a machine that matters, the default changed underneath you today.
✓ Verified · 3 sources
▶ Related video: How auto mode works with Claude Code
Read in the app — free, in 9 languages
Related stories
OpenRouter dropped three AI providers after a bait-prompt leak test.
2026-11-22Researchers say they broke out of Meta's Muse agent and leaked chats.
2026-11-22ChatGPT's shared workspace Space is now rolling out to Plus users.
2026-11-21Amazon says attackers' AI agents found an unknown flaw on their own.
2026-11-21METR says it can rarely find humans who beat top AI at coding tasks.
2026-11-21