Anthropic makes Claude Code Auto Mode the default from August 14

Anthropic’s Claude Code will switch to Auto Mode by default on August 14 for Pro, Max and Team plans, a change the company says is designed to protect developers from careless manual approvals. Auto Mode lets the AI coding assistant run longer stretches of work on its own and stops only to ask for confirmation when a classifier decides an action is dangerous or irreversible. Enterprise customers will still need to opt in.

What becomes the default on August 14
Auto Mode is not a new feature; it has been available in Claude Code for some time. What changes is the default behaviour. Unless a user changes the setting, Pro, Max and Team subscribers will run in Auto Mode from August 14, while Enterprise administrators can decide whether to keep manual approval flows. Anthropic said in a blog post that Auto Mode adds a protective layer against prompt injection attacks, where hidden instructions embedded in code try to hijack the model away from the user’s actual task. The company also confirmed it will not charge for the tokens that the classifier itself consumes while deciding whether an action needs human approval.

Por que importa

How the safety numbers stack up
The central claim is that the classifier catches risky commands more reliably than humans do. According to Anthropic’s tests with 1,053 paid testers and internal red-teaming, the classifier flagged 89 percent of dangerous commands, while human reviewers caught only 13.6 percent when judging approval requests manually. The company says Auto Mode performed at least as safely as manual approvals, and in several cases better. Teams using Auto Mode also produced about 25 percent more pull requests during the tests, which Anthropic reads as a sign that developers got more useful work done without sacrificing safety.

Prompt injection tests: 720 attempts, zero successes
An independent audit by Trajectory Labs tested 72 attack scenarios against Claude’s current models — Fable 5, Opus 5 and Sonnet 5 — with each scenario repeated ten times in Auto Mode. None of the 720 attempts succeeded. The same style of testing against OpenAI’s GPT-5.6 Sol in Codex Auto-Review mode let 5.83 percent of attacks through. Anthropic also shared internal incidents. In one session, Auto Mode stopped Claude from uploading confidential data to a public page. In another long-running session, the company says it killed roughly 2,000 processes that would have disrupted active GPU training jobs.

Contexto

The oversight paradox
Claude Code is currently the most widely used AI coding tool by a large margin, so this default change affects a big slice of the developer population. The practical effect is that developers shift further from writing code to supervising AI output. Anthropic itself urges caution: “For high-stakes changes to production infrastructure, we still recommend reviewing Claude’s actions yourself.” That advice creates a tension. When a tool asks for approval less often, the approvals that do happen carry more weight, yet developers have less direct experience of the codebase because much of it was written by the model without them. The result is an oversight gap that tools alone may not close.

What remains unclear
The safety figures come from Anthropic’s own test design; they are not an independent benchmark of day-to-day coding safety, and the 89 percent figure says nothing about the commands the classifier might still miss. The Trajectory Labs audit covers prompt injection only, not broader risks such as data exposure or unwanted code changes. Anthropic acknowledges that the classifier reduces risk but does not eliminate it. There is also a commercial angle: making Auto Mode the default lets Claude work longer and consume more tokens, which raises usage-based revenue even if that was not the stated motivation. What remains unmeasured is how Auto Mode affects long-term code quality, maintenance burden, and senior developers’ ability to review unfamiliar code written largely by an agent.

| Fact | Detail | Source |
| What changes | Auto Mode default for Pro, Max, Team from August 14; Enterprise opts in | Anthropic via The Decoder |
| Test scale | 1,053 paid testers plus internal red-teaming | Anthropic tests |
| Dangerous command detection | Classifier 89% vs human reviewers 13.6% | Anthropic via The Decoder |
| Prompt injection audit | 0 of 720 attempts succeeded against Claude; 5.83% against GPT-5.6 Sol | Trajectory Labs via The Decoder |
| Productivity signal | About 25% more pull requests in Auto Mode teams | Anthropic tests |

Source: The Decoder — https://the-decoder.com/anthropic-sets-claude-code-to-auto-mode-by-default-to-protect-developers-from-bad-approvals/

Datos clave

Punto	Detalle
Fuente	The Decoder
Fecha	2026-08-08T14:58:57+00:00
Tema	Anthropic sets Claude Code to Auto Mode by default to protect developers from bad approvals