Anthropic: AI classifier replaces human approval in Claude Code

Anthropic is changing the safety logic of its developer tool Claude Code: from 14 August 2026, automatic mode becomes the default for the Pro, Max and Team plans. Instead of a human manually approving every potentially risky command, an AI classifier takes over the review.

The trigger was a sobering study with 1,053 testers. The result: humans detected dangerous commands only 13.6 percent of the time. The AI classifier reached 89 percent. Even more striking: after roughly 50 approvals in a row, the human hit rate dropped to about 5 percent – a classic case of 'click fatigue', where users wave warnings through reflexively.

The switch is a remarkable admission: for repetitive safety decisions, the human is not the better guardian but the weakest link. The AI does not tire, does not get sloppy, and judges the thousandth command as carefully as the first.

For the growing number of companies deploying coding agents and AI-driven automation, this is highly relevant. It shifts the human's role: away from nodding through individual actions, toward setting rules, sample auditing, and stepping in on exceptions.

At the same time, the fair counter-question remains: who watches the watcher? Anthropic bets that a specialized, transparently evaluated classifier is more reliable than overtired humans. The study supports the thesis – but the real test comes in the daily flow of millions of automated commands.