Humans approving AI agent commands missed 1 in 3 threats across 40,000 runs

Data from over 40,000 plays of a browser game simulating human-in-the-loop approval for an AI coding agent shows the average player missed 1 in 3 threats, with mean accuracy of 66.3% across 409,000 approve/deny decisions. Nearly a third of sessions ended with a negative score, and 7% approved every single prompt. Only 20.8% of players caught every threat while also keeping false blocks under 1 in 5; the rest achieved perfect detection largely by blocking everything. Obvious destructive commands like `rm -rf /` were missed 11.7% of the time. The caveat: 34% of commands in the game were threats, far above real-world base rates, where vigilance decays further.

View full digest for August 7, 2026