A large-scale evaluation of human reviewers approving AI agent commands across 40,000 runs found that humans missed one in three security threats. The study also revealed that reviewers frequently blocked safe commands, demonstrating significant flaws in current human oversight processes.