Anthropic makes Claude's Auto Mode the default to cut approval clicks

What's going on

On October 22 (local time), Anthropic updated its announcement: starting next week, "Auto Mode" becomes the default in most Claude apps for newly started sessions on the Pro, Max and Team plans. Until now, you often had to confirm actions such as running tools one by one. Now an automatic check reviews every request: actions classified as safe go ahead, and risky ones are blocked. That should let Claude work for longer stretches without interruption.

How Auto Mode works

Anthropic already made the mode the default in August in Claude Code, its coding tool. Now it's being extended to other Claude apps. If an action is blocked, Claude can look for another route or ask you for permission. If you prefer, you can switch back to manual approval. For enterprise plans, Anthropic notifies administrators in advance via email.

The checks cover things like changes that are hard to reverse, destructive commands and actions that send data outside. As examples from internal use, Anthropic cites blocked attempts to publish a report on a public code-sharing site, and an action that tried to grant cloud admin rights beyond what was requested.

Why Anthropic is cutting the clicks

The company argues that constant confirmations offer little protection. Among Claude Code users, 97 percent of approval requests were accepted. Most involve routine tasks, but if you're asked dozens or hundreds of times a day, you often click out of habit without looking.

The test numbers

The study was run as a programming experiment with 1,053 paid professional testers, who had dangerous commands slipped into approval requests. Important: the figures come from Claude Code. That they hold equally for all the apps now affected isn't proven. Anthropic says insights from internal tests and external security assessments have fed into improving the mode across all products. All figures are company claims.

Our take

The move shifts safety from human attention to automated review, which is practical for long agent tasks. The data suggests that constant prompts protect poorly. At the same time, Anthropic admits that even Auto Mode doesn't eliminate risk entirely. For major changes to production infrastructure, the company recommends reviewing them yourself. Independent verification of the numbers is still pending, and so is the question of how often harmless actions are wrongly blocked. The source says nothing on that.