Anthropic Makes Claude Code Auto Mode Default, Outperforming Manual Review
机器之心 · wechat · 2026-08-08
Anthropic announced that starting August 14, Auto Mode will become the default permission mode for Pro, Max, and Team users in Claude Code. The mode uses a separate classifier to review shell commands, successfully intercepting 89% of dangerous commands in tests, significantly outperforming the 14% caught by manual human review.
- Limitations of Manual Review: Data shows users blindly approved 97% of permission prompts, making manual oversight ineffective. Auto Mode not only enhances safety but also allows tasks to run unsupervised in the background for hours. Teams using Auto Mode saw a 25% increase in submitted PRs.
- Security & Defense: The Auto Mode classifier adds a defense layer against prompt injection attacks. In independent evaluations, all 720 attack attempts against Claude models failed; in contrast, GPT-5.6 Sol running in Codex's auto-review mode had an attack success rate of 5.83%.
- Interception Instances: The classifier successfully blocked outbound data exfiltration, destructive mass process terminations targeting training GPUs, and excessive cloud privilege escalations.
Additionally, Anthropic will no longer charge these subscription tiers for the extra tokens consumed by the Auto Mode classifier.
Related event: Anthropic Sets Claude Code Auto Mode as Default on August 14(5 posts)→
More from coding & agent
- TensorLens: Inspect HF Model Quantization Layouts Directly in Your Browser — Brilliant-Hall1387 · 2026-08-08
- Claude Plays Doom: Tracking AI Decisions Frame-by-Frame with W&B — _ScottCondron · 2026-08-08
- Dev Discussion: What AI Agent Skills Are Actually Useful in Daily Coding? — jonathan_wilke · 2026-08-08
- Alibaba Open-Sources Page Agent: A Pure JavaScript In-Page GUI Agent — thisguyknowsai · 2026-08-08
- AI Alignment Getting Easier, But Unrestricted Agents Pose Security Risks — teortaxesTex · 2026-08-08
- Vercel Open-Sources Knowledge Agent Template Using grep Instead of Vector DBs — tom_doerr · 2026-08-08