Research · curated 6 Aug 2026
Humans in the loop miss a third of dangerous AI coding agent requests
First reported scalex.dev
Coverage timeline
Single-source research — first reported, latest, and curated coincide.
Why it matters
Human-in-the-loop approval is a primary safeguard against AI coding agents running destructive or data-exfiltrating commands, and these results show it fails at a high rate under fatigue — meaning defenders cannot rely on manual approval alone to contain agentic blast radius.
A browser-based game built by developer Alex Wauters tested humans' ability to approve or deny AI coding-agent permission requests under time pressure, and across 40,000+ runs and 409,000 decisions players let roughly one in three malicious commands through. Scope violations like an agent trying to cat AWS credentials or Kubernetes config were missed 35% of the time, and 'npm run analyze' slipped by nearly 65% of the time; Anthropic's own telemetry separately showed users approved about 93% of Claude Code permission prompts, reflecting approval fatigue.