AI agent · AI securityOther industries (software / AI assistants)
Anthropic is making automated action-checking the default in most Claude apps, replacing per-action human approval.
Source [1] AI타임스 · 23 Oct 2026 · “앤트로픽, 클로드 ‘오토 모드’ 기본 적용…사용자 승인 절차 줄인다” ↗
Adoption stage: Production — A named organisation runs it in day-to-day operations.
Summary
Anthropic is making automated action-checking the default in most Claude apps, replacing per-action human approval. Its own Claude Code data shows people approve 97% of prompts and catch only 13.6% of planted dangerous commands, falling to about 5% after 50 approvals, against 89% for the automated checker. The numbers are from the coding tool, not yet from the wider apps.
Kenny's take
Anyone who has signed the same deck checklist forty times knows the fiftieth signature is a reflex, not a check. The software world now has numbers for it. Keep the human for the irreversible calls — a dangerous-goods restow, a sailing decision — and let an automated gate screen the routine. If the 89% does not hold outside coding, this take is wrong.
AI draft · approved by Kenny
Analysis
This is a named vendor moving an "automated reviewer in front of agent actions" pattern into day-to-day use for its own users — production for that pattern in software. More important for the lens is the failure-mode data on human-in-the-loop: the approval gate degrades with volume (13.6% → ~5% after 50 prompts). That directly challenges a naive reading of thesis 1 (human as final authority) and supports thesis 3 (harness matters more than the model).
Inference: operational HITL designs in maritime/logistics that put a human click on every agent action will show the same fatigue curve; the gap between software (production) and operations (approval gates still designed per action) is a design gap, not a model gap. What the article does not show: how auto mode performs outside coding, false-block rates, or any operational-technology setting.
Facts from the source
- F1Anthropic said on the 22nd (local time) that from next week "Auto Mode" will be the default for new sessions in most Claude apps for Pro, Max and Team plan users; enterprise plan administrators will be notified by email before the default applies. [1]
- F2The approach extends what Anthropic made the default in its coding tool Claude Code in August. [1]
- F3In auto mode, tool-execution requests are checked automatically: actions judged safe proceed, actions judged risky are blocked, after which Claude may find another route or ask the user for approval; users can switch back to manual approval. [1]
- F4Anthropic's analysis found Claude Code users approved 97% of approval requests. [1]
- F5In a coding experiment with 1,053 paid professional testers, humans blocked 13.6% of clearly dangerous commands inserted into approval requests; auto mode blocked 89% of the same commands. [1]
- F6When a person had already handled 50 or more approval requests, the human block rate fell to about 5%, while auto mode's block rate did not vary with session length. [1]
- F7The figures come from Claude Code tests, not from the general apps now getting the default; Anthropic recommends users still review significant changes to production infrastructure themselves. [1]
Organisations: Anthropic
Counterpoint — verification notes
Codex checked each sentence against the source and searched for counter-evidence. Verdict: Codex check · partly supported.
- Other sideAnthropic's October 22 update confirms the planned rollout in most apps. Its guidance also says users with a different default may be prompted to switch and a pinned default remains unchanged, further limiting an unqualified claim that the default applies across Claude apps. [2]
2 sentence-level checks
- summary_kopartly-supported: 앤트로픽이 클로드 앱 전반에 '오토 모드'를 기본값으로 적용한다. — “앱 전반에” drops the limits to most apps, eligible plans and new sessions. The wider rollout was announced for the following week.
- summary_kopartly-supported: The 97% approval rate and 13.6% versus 89% block rates are all figures from Anthropic's own experiment. — The 97% figure came from Claude Code usage analysis, while 13.6% and 89% came from the coding experiment. Unlike the English summary, the Korean summary also omits the coding-only scope and the approximately 5% result after 50 prior approvals.
Sources
- [1] AI타임스 — “앤트로픽, 클로드 ‘오토 모드’ 기본 적용…사용자 승인 절차 줄인다” original
- [2] claude.com counter-evidence (Codex)
Keep reading
Stay in the loop.