Human Supervisors Miss 1 in 3 Security Threats From AI Agents

Placing a human in the loop to review AI agent actions isn't the safety net we hoped for.

A browser game simulating a human-in-the-loop supervisor for AI coding agents tracked over 40,000 runs and 409,000 approval decisions. According to published stats from the creator, players missed 1 in 3 security threats, posting an average accuracy of 66.3%. Obvious destruction like rm -rf / was caught 88.3% of the time, but exfiltration hidden inside familiar script triggers like npm run analyze slipped past 64.7% of the time.

Why it matters: Asking users to approve individual shell commands creates a false sense of security. An agent can plant malicious payload code in package.json before triggering a standard build step, making dangerous actions look completely innocent. Add permission fatigue to the mix, and human supervision quickly turns into blind rubber-stamping.

Know this: Players also over-blocked safe commands out of paranoia, rejecting harmless cleanup tasks like rm -rf dist/ 45% of the time. This friction slows work down, causing users to drop their guard even further.

At least 7% of players approved every single prompt without looking—true fans of living dangerously.