Anthropic Makes Claude Code Autonomous by Default
From 14 August, Claude Code stops asking permission at each step. The justification — users approve 97% of permission prompts — is worth taking to your own approval workflows.
From 14 August 2026, Anthropic switches Claude Code's auto mode on by default for Pro, Max and Team accounts. Auto mode lets the agent run without asking permission at each step. It stops to ask only when an action looks irreversible, destructive, or aimed outside your environment.
Auto mode isn't new — it was unveiled in March 2026. What changes is the default, and defaults are the whole argument.
The data Anthropic used to justify it
Anthropic tested with 1,053 paid users and reported a result that should make anyone who has designed an approval workflow uncomfortable: auto mode caught 89% of harmful actions, while manual human review caught 13.6%.
The explanation is in a single statistic. Users approve 97% of permission prompts. At that rate the prompt is not a control. It's a speed bump that people have learned to drive over, and it manufactures a paper trail suggesting a human checked something they did not check.
Boris Cherny, who heads Claude Code, said his team has used auto mode exclusively for months.
What replaces the prompt
Two things. Prompt injection screening, so instructions smuggled in through files or fetched pages get scrutinised rather than executed. And customisable hard deny rules, so you can name the things the agent must never do regardless of what it concludes in the moment.
That's a meaningful architectural shift: from per-action consent to policy set in advance. It is also, notably, how we already govern most automation in the enterprise. Nobody approves each row a batch job writes.
What I'd do about it before the 14th
If you run Claude Code inside an organisation, the deny rules are now the security boundary and they need to be written by someone who knows what your credentials can reach. Default deny rules are generic; your blast radius is not.
The 97% approval figure deserves to travel beyond this one product. Look at your own change-approval gates, your access-request queues, your model-deployment sign-offs. If the approval rate is in the high nineties, you don't have a control there either. You have a ritual, and swapping it for a written policy would be an improvement rather than a compromise.
Source: Anthropic is turning Claude Code's auto mode on by default — TechCrunch