Humans in the loop miss a third of dangerous AI coding agent requests
You wouldn't let Claude Code cat your AWS credentials or Kubernetes config on request, would you?
A new browser-based game designed to test human ability to safely approve AI coding agent requests reveals that humans in the loop are missing a third of dangerous commands on average. The game, created by Belgian software developer Alex Wauters, simulates permissions requests like those users might encounter from AI agents during workflow execution.
Players are given 60 seconds to approve or deny as many requests as possible, with both approved malicious and denied safe commands subtracting from their score. The results, based on over 40,000 runs, show that humans in the loop fail to catch roughly one in three malicious requests, with scope violations being the most commonly missed.
Additionally, the game highlights the potential for fatigue and decreased vigilance when manually approving an agent's actions, as well as the risk of approving destructive commands under time pressure.
Brief written by urgent.news from The Register Software's own syndicated text. Machine-written — it may contain errors, so check the original before relying on it.
