Claude Code Makes Auto Mode the Default for Pro, Max and Team
From August 14, new Claude Code sessions on Pro, Max and Team plans start in auto mode, where a classifier instead of the user reviews each tool call before it runs.
TL;DR
- From August 14, new Claude Code sessions on Pro, Max and Team plans start in auto mode, where a classifier checks each tool call [^1].
- Anthropic argues that prompts have become reflexive: in its study, testers caught 13.6 percent of planted dangerous commands, the classifier 89 percent [^1].
Background
Claude Code is Anthropic's coding agent: it reads a codebase, edits files and runs shell commands. By default it has asked the user to approve actions through permission prompts. Users can widen what runs without a prompt through allow rules in their settings, such as one that lets any Python command through, or skip prompts with a mode called bypassPermissions [^1]. Auto mode, which was already available as an option, replaces the prompts with a classifier [^1].
What happened
Anthropic announced on August 7 that new sessions on its Pro, Max and Team plans will run in auto mode starting August 14 [^1]. Users who set a different default may get a one-time prompt asking whether to switch, and a pinned default stays as it is. The classifier uses a few extra tokens per tool call, and Anthropic stopped charging Pro, Max, Ultra and Team users for that overhead [^1]. Auto mode stays opt-in for now on Claude Enterprise, the Claude API, Amazon Bedrock, Microsoft Foundry and other cloud platforms; Anthropic plans to make it the default there in the coming month [^1].
In auto mode, every tool call goes through a classifier aimed at blocking actions that are irreversible, destructive or aimed outside the user's environment [^1]. When it blocks something, Claude usually looks for a safer route or asks the user. After three blocks in a row, or twenty in a session, Claude Code falls back to manual approvals [^1]. Allow rules broad enough to grant arbitrary code execution, such as python:*, are set aside in auto mode so that commands can't skip the classifier. Sending code or secrets to an external destination is a hard deny the classifier is designed never to approve, and before a git push the classifier checks whether the destination is public, private or trusted [^1].
Anthropic backs the change with its own data. Users approve 97 percent of permission prompts, while they reject 39 percent of the plans Claude presents [^1]. In a study with 1,053 paid testers, a single prompt per session was swapped for a clearly dangerous command. Testers caught it 143 times, or 13.6 percent; auto mode blocked 937 of the same commands, or 89 percent [^1]. Human catch rates fell from about 17 percent early in a session to about 5 percent after 50 or more prompts. In flagged real sessions from May to June, 6.3 percent of manually approved sessions contained a production-level harmful action the user hadn't explicitly asked for, against 2.4 percent in auto mode [^1].
Why it matters
The change moves the default safety check from a person to a model. Anthropic's settings data suggests that for many users the alternative was not careful review: as of June, 49.5 percent of active CLI users had created a Bash allow rule, and 25 percent of interactive sessions started in bypass permissions mode [^1]. Measured against that habit, a classifier that sees every call can be the stricter option.
The evidence comes from Anthropic and from firms it hired, and Anthropic adds its own caveats. In red-teaming by Apollo Research, the classifier's miss rate on synthetic attacks fell from 12 to 7 percent after hardening, a figure Anthropic says should not be read as the miss rate on real traffic [^1]. The company also says auto mode relies on classification and does not eliminate risk, and it still recommends reviewing Claude's actions yourself for high-stakes changes to production infrastructure [^1].
Practical example
Say you use Claude Code on a Pro plan and never set a default mode. After August 14, a new session shows a notice and starts in auto mode [^1]. Your old allow rule for python:* stops applying while you're in auto mode, so the classifier now looks at those commands too; the rule returns as soon as you switch modes, and your settings file isn't touched [^1]. When the classifier blocks a step, you'll see Claude try another route or ask you. If you'd rather approve everything by hand, Shift+Tab in the CLI switches modes [^1]. Before a session that touches production servers, you go back to reviewing each action, as Anthropic suggests [^1].
Related gear
We recommend this book because it covers habits such as version control, testing and automation that help you review code changes, whoever wrote them.