Claude Code makes Auto mode the default Aug 14 for Pro, Max, Team

A classifier now vets every command by default — 89% of dangerous ones caught to humans' 14%, and 720 prompt injections stopped cold. Opt-out is one keystroke.

Nowline AUG 9 10:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • The default flips on August 14

    Starting Aug 14, Auto mode becomes the default permission mode for new sessions on Pro, Max, and Team plans; Enterprise and the API stay opt-in, with their rollout due within a month. Instead of approving each step, a classifier gates every tool call and prompts you only when it flags something irreversible, destructive, or aimed outside your environment.

  • 89% caught, versus 14% for humans

    In a controlled trial with 1,053 paid developers, Auto mode's classifier blocked 89% of dangerous commands while manual click-through review caught just 13.6%. Head to head it stopped 800 commands a human had approved, and humans blocked only 6 that Auto mode let run — the pitch is that approving prompts all day is exactly where you get sloppy.

  • Zero of 720 prompt injections landed

    A third-party eval by Trajectory Labs fired 720 indirect prompt-injection attempts across 72 scenarios; none got through against Claude Fable 5, Opus 5, or Sonnet 5 in Auto mode (GPT-5.6 Sol let 5.83% through in Codex's auto-review). The gap it doesn't close: a malicious npm or PyPI package you install yourself can still exfiltrate data, so vet dependencies as carefully as ever.

  • Keeping a hand on the wheel

    Switch modes anytime with Shift+Tab or the desktop dropdown, and a default you set yourself sticks unless you accept the one-time prompt. For hard checkpoints, add ask rules like `Bash(git push *)` and `Bash(gh pr create *)`; teach the classifier your infra with `autoMode.environment`; and run `claude auto-mode reset` to wipe custom rules.

  • Elsewhere: the US DOE ships an open science model

    The Department of Energy launched the Genesis Open Models Initiative with Arcee AI, whose first model — Genesis-Science-1 — is trained on real national-lab work: HPC code, simulation campaigns, materials science, and energy systems. The release will bundle open weights, a technical report, and public workbench demos; contributor applications close Aug 14.