A small browser game that simulated a human-in-the-loop for an AI coding agent collected data from over 40,000 runs and 409,000 individual decisions. The results showed that only 35.2 percent of players caught every threat, while just 20.
log in to read full article