Humans missed 1 in 3 threats while approving AI agent commands in a browser game involving 40,000 plays. The game simulated a human-in-the-loop scenario where players approved or denied commands under time pressure, revealing vulnerabilities in command approval processes.