Humans missed 1 in 3 threats approving AI agent commands across 40k game runs

A browser-based game testing human oversight of AI coding agents revealed that players missed approximately 1 in 3 malicious commands, with an average accuracy of 66.3% across over 40,000 game runs. Credential exfiltration and scope violation threats were missed most frequently (33-35% miss rates), while obviously destructive commands like `rm -rf /` were caught more reliably (11.7% miss rate). The most-missed command was `npm run analyze`, with two-thirds of players approving it despite visible evidence in the history log that it contained malicious code, suggesting players often fail to carefully review agent action details before approval.