Anthropic puts Claude Code on autopilot by default from August 14
The AI coding tool will now act on its own and only pause for truly risky steps. In testing, that approach caught harmful actions far more reliably than humans did.
Key points
- Anthropic will make auto mode the default for Claude Code's Pro, Max, and Team accounts starting August 14, 2026.
- In a study of 1,053 paid testers, auto mode caught 89% of harmful actions; human reviewers caught only 13.6%.
- Humans approved 97% of permission prompts during manual review, suggesting routine click-through had made oversight nearly meaningless.
- Anthropic is also adding prompt injection screening and customisable hard deny rules to Claude Code.
Anthropists have long assumed that keeping a human in the loop makes AI tools safer. A new study from Anthropic suggests the opposite may sometimes be true, at least when the human in question has seen the same approval box a hundred times today.
Starting August 14, Claude Code, Anthropic's AI-powered coding assistant, will switch on auto mode by default for users on Pro, Max, and Team plans. Auto mode means the tool completes tasks without stopping to ask permission at every step. It pauses only when an action is judged irreversible or destructive, or when it reaches outside your own environment.
Anthropic first tested auto mode in March 2026, pitching it as a way to keep work moving without sacrificing control. Our coverage of Claude Code goes back to 20 July 2026, and the pattern this announcement fits is one we flagged as early as 6 August: a browser game simulating real AI coding requests found players approved roughly a third of malicious commands across more than 40,000 runs.
Why does removing human approval actually make things safer?
Because humans stop paying attention. In Anthropic's study, users approved 97% of all permission prompts during manual review. Clicking "yes" had become muscle memory, not a genuine check.
Auto mode applies consistent rules every single time. Across 1,053 paid testers, it correctly flagged 89% of harmful actions. Human reviewers caught just 13.6% of the same actions.
Boris Cherny, head of Claude Code, put it plainly in a post on X: "The team and I use Auto mode exclusively, and have been for many months. I couldn't imagine going back to permission prompts."
What should users actually watch for?
Auto mode is not a free pass to walk away while Claude Code rewrites your codebase.
The tool pauses before actions it judges harmful, but that judgment is made by the AI, not by you. Getting familiar with the new hard deny rules, which let you block specific categories of action outright, is worth the ten minutes it takes.
Anthropic is also rolling out prompt injection screening. Prompt injection is an attack where malicious instructions are hidden inside content the AI reads, such as a webpage or a document, tricking it into acting against your wishes. The screening is designed to catch those hidden commands before they change what the tool does.
If your work touches production systems that can't easily be rolled back, set your hard deny rules before August 14. Auto mode is a capable assistant that works fast and needs clear boundaries. Not one to ignore entirely.
