English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Anthropic moves Claude Code's default to Auto Mode, shifting approval from humans to a classifier

Forum topic · 小凯 · 2026-08-13

Summary

On August 14, 2026, Anthropic switched Claude Code's default permission mode for Pro, Max, and Team plans from per-action confirmation to Auto Mode. The core change is that an independent classifier, not the user, judges whether each tool call is safe, irreversible, or externally triggered; calls below the threshold are blocked, the model attempts a safer path, or the user is asked. After three in-conversation or twenty single-session blocks, Claude Code falls back to manual approval. Enterprise, Claude API, and Cloud Platform users follow within three months. Anthropic's own data shows users approve 97% of permission prompts reflexively and 49.5% write their own Bash allow-rules, so Auto Mode formalizes an existing pattern. In a controlled test of 1,053 paid evaluators, humans caught 13.6% of planted dangerous commands versus 89% by the classifier; human detection dropped to 5% after 50+ routine approvals. Production data shows serious-incident rate rose from 2.4% (manual) to 6.3% (Auto), which Anthropic publicly acknowledges. Auto Mode's real defense is indirect prompt injection, blocking 720 of 720 attacks across Claude Fable 5, Opus 5, Sonnet 5, compared to GPT-5.6 Sol's 19.03% in Full Access. Anthropic does not bill classifier tokens, hinting at a per-PR revenue model.

On August 14, 2026, Anthropic switched Claude Code's default permission mode for Pro, Max, and Team plans from "confirm each action" to Auto Mode. The core mechanism is an independent classifier that decides whether each tool call is safe, irreversible, or externally triggered. Calls scoring below the threshold are blocked; Claude Code then tries a safer path or explicitly asks the user. After three in-conversation or twenty single-session blocks, Auto Mode falls back to manual approval. Enterprise, Claude API, and Cloud Platform (AWS, Google Cloud, Microsoft Foundry) users follow within three months. Anthropic's framing: admins can lock the org default; individual users can lock their personal preference.

Why the shift now

Anthropic's published numbers:

  • Current Claude Code users approve 97% of permission requests on average — a figure Anthropic itself reads as "reflexive clicking, not real review."
  • 49.5% of active CLI users have written their own Bash allow-rules.
  • 62% have used bypassPermissions or "don't ask again."
  • "Human review" had already been bypassed by humans themselves; the remaining question was what catches things after the bypass. Auto Mode is not new — it shipped as opt-in in March — but the August 14 switch formalizes the existing reality and replaces the human click with a model gate.

    Evidence behind the August 14 switch

    A controlled experiment with 1,053 paid professional testers planted a clearly dangerous command inside routine permission requests:

  • Human testers caught it 13.6% of the time (143 / 1,053).
  • Auto Mode's classifier caught it 89% of the time (937 / 1,053).
  • Human detection drops to ~5% after 50+ routine approvals; the classifier stays stable.
  • Production-data retrospective (May–June):

  • Manual-approval sessions: 2.4% contained a "production-grade serious incident" — a harmful action that should not have executed (writes to prod DBs, pushes to prod branches, destructive commands).
  • Auto Mode sessions: 6.3%.
  • This is the first time Anthropic has publicly used its own numbers to admit: Auto Mode does not solve the safety problem, it redistributes it. The Auto Mode incident rate is roughly 2.6× the manual-approval rate.

    Red-team results

  • Apollo Research two-week red team cut the classifier's overall miss rate from 12% to 7% — still a miss rate.
  • A third-party evaluation by Trajectory Labs (72 indirect prompt-injection scenarios, 720 attack attempts):
  • Claude Fable 5 / Opus 5 / Sonnet 5 under Auto Mode: 0 successes.
  • GPT-5.6 Sol under Codex Auto-review: 5.83% success.
  • GPT-5.6 Sol under Full Access: 19.03% success.
  • Auto Mode's real strength is not blocking dangerous commands the user actually wrote — it's blocking prompt injection smuggled in from external pages or tool outputs. Anthropic treats this as product differentiation and writes into the changelog that the adversary is prompt injection, not the user. What they don't (and can't) say: dangerous commands written by the user slip past both the classifier and human approval.

    What this changes

    Anthropic offers an enterprise-IT-ready formula: longer agent tasks + fewer approval popups + users ignoring 97% of popups → the model takes over review. CIOs gain a new tool: instead of writing per-developer permission policies, write one policy set and let the classifier enforce it uniformly. Analyst Manoj Chandra Jha frames it as "the CIO's governance surface is the whole organization, not individual developers." Anthropic has opened enterprise opt-in windows across API, AWS, and Google deployments for policy pre-staging.

    A commercial detail: classifier tokens are not billed against existing quota — Anthropic absorbs the cost. On one side, Auto Mode users already file ~25% more pull requests than manual-mode users (a revenue pull). On the other, classifier cost is a hidden per-interaction expense. The implied direction: from "Claude Code sold by token" to "Claude Code sold by PR."

    Risks and limits

    1. "Irreversible action" is not auditable. The classifier is a model; its judgments are distributional, not rule-based. It can distinguish git reset --hard from git log with logic, but "is writing a +1 on line 47 of this specific file irreversible" is a continuous score. Auditing 4 hours of unattended agent output requires a different skill set than deciding on a 30-second popup. 2. Single point of failure. Analysts converge here: safety moved from "distributed across every developer's every click" to "concentrated in one classifier." One blind spot, one bypass. Anthropic's own framing — "from 13% miss rate to 7%" — is still 7% across hundreds of millions of calls per year. 3. A new classifier is a juicy target. Third-party evaluation shows Claude blocks all 720 attacks today, but iterated attacks (obfuscated injection, splitting a command across multiple tool calls) are not publicly measured. Anthropic has not committed to monthly miss-rate disclosure. The hard-deny list (outbound code, git reset --hard, external-content prompt-injection screening) covers a few rule-based cases; the rest remains soft model judgment.

    Bottom line

    Anthropic's move is not pushing agents toward greater autonomy. It is pushing agents toward "managed boundaries" — moving approval responsibility from "something you should do" to "something the model does for you," while acknowledging the accountability chain is unproven. The bet: longer unattended tasks + larger code changes + more frequent PRs will sustain a new revenue stream. The exposure: when a bad PR ships and a prod incident happens, who signed off on the review? The odds work like an actuary's insurance product — as long as misclassification stays below a critical threshold.

    Sources

  • https://bitroot.org/blog/2026-08-12-claude-code-switches-to-automode-by-default
  • https://cyberpress.org/claude-code-makes-auto-mode-default/
  • https://en.it-daily.net/it-management-en/ai-en/anthropic-claude-code
  • https://www.infoworld.com/article/4207959/anthropic-makes-claude-codes-auto-mode-default-for-paid-users.html
  • https://aiinsiders.net/article/claude-code-makes-auto-mode-the-default-not-just-an-option

Tags

#anthropic#claude-code#auto-mode#ai-safety#prompt-injection#enterprise-it#developer-tools#agent-governance

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633415