Claude Code will stop asking permission by default from 14 August — because humans catch 14% of dangerous commands and the classifier catches 89%
From 14 August 2026, Anthropic is switching Claude Code's default from repeated approval prompts to 'auto mode', a classifier that screens tool calls for irreversible or destructive actions — a change Anthropic says is justified by testing showing human reviewers catch only 13.6% of dangerous commands against the classifier's 89%.
8 August 2026
Anthropic has announced that Claude Code’s “auto mode” becomes the default permission setting from 14 August 2026 for Pro, Max and Team users, replacing the familiar pattern of approving each tool call individually. In its place, a classifier screens every action Claude Code wants to take for signs it’s irreversible, destructive, or outside the task’s scope, and only interrupts the workflow for those.
The justification is a specific number, not a vague appeal to convenience: in a study of 1,053 paid testers, human reviewers correctly caught only 13.6% of genuinely dangerous commands when asked to approve or reject them — a figure that fell to roughly 5% after the reviewer had already approved 50 prior prompts, the classic pattern of approval fatigue. The auto mode classifier caught 89% of the same dangerous commands in testing. Anthropic is also dropping the token charge for the extra classifier calls, removing the one obvious reason a team might have stayed on manual approval to save cost.
Two things are worth noting alongside the headline change. First, Anthropic is explicit that “classifiers cannot eliminate risk” and still recommends human review before anything reaches production — this is a change to the default review mechanism, not a claim that review is no longer needed. Second, OpenAI has taken the opposite position with GPT-5.6, opting out of an equivalent auto-approval default — a genuine philosophical split between the two labs on how much autonomous execution risk is acceptable by default, not just a feature gap.
So what
If your team runs Claude Code today, this changes your risk posture automatically on 14 August unless someone actively pins the previous manual-approval setting — worth a deliberate decision either way rather than letting the default choose for you. And the underlying data point matters beyond this one product change: it’s a rare, quantified admission that manual human approval of AI-agent actions is measurably worse than it feels, which should inform how any business scopes permissions and guardrails around AI coding tools generally, not just this one setting in this one product. If you want help thinking through where automated AI development genuinely needs a human in the loop and where it doesn’t, see our approach to AI-assisted development or get in touch.