A browser-based game simulating AI coding agent permission requests found that human reviewers approved roughly one in three malicious commands, with fatigue from repeated approvals being a key factor — a finding echoed by Anthropic's own telemetry showing users approve around 93% of all permission prompts. The most commonly missed threats were subtle scope violations and ambiguous commands like `npm run analyze`, while obviously destructive actions were more reliably caught. Experts argue that simply relying on humans as a last line of defense is insufficient, and that safer solutions require sandboxing, automated classifiers, and better-designed permission models for AI coding agents.
