August 6, 2026
Click Yes, Regret Later
Humans missed 1 in 3 threats approving AI agent commands across 40k game runs
People smashed the AI approve button and let 1 in 3 dangerous moves slip through
TLDR: A test game found people approved dangerous AI actions surprisingly often, especially when the command looked familiar and harmless. In the comments, many said this proves popup-based safety is a joke, while others argued humans still help if used sparingly and backed by better safeguards.
The real tea here is not just that people missed one out of every three dangerous AI actions in a browser game built to test human oversight — it’s that the comment section reacted like this was the least surprising plot twist on Earth. Across more than 40,000 game runs, players were asked to approve or reject an AI assistant’s commands under pressure. They were great at spotting the cartoon-villain stuff, like obvious destructive commands, but far worse at catching the sneaky moves that could quietly expose private account keys or send data away. And the biggest facepalm? Familiar-looking commands like "npm run analyze" got waved through most of the time, even when the warning signs were sitting right there in the history log.
That sparked the hottest reaction on Hacker News: why are we still pretending endless permission popups are security? One commenter basically laughed at the whole model, saying software keeps trying the same failed trick — ask users constantly and pray they don’t mess up. Another went even sharper, calling the whole thing a legal cover move so companies can shrug and say, well, you clicked yes. Ouch. Still, not everyone was in full doom mode: one camp argued that 1 in 3 isn’t catastrophic if humans are used more selectively, backed by extra filters and warning systems. So the community split into two loud factions: the “this system is broken and silly” crowd, and the “fine, but only if humans aren’t forced to judge every tiny action” optimists. The vibe was equal parts cautionary tale, workplace comedy, and giant flashing sign that people will absolutely click the friendly-looking button.
Key Points
- •The browser game collected data from over 40,000 runs and 409,000 approve/deny decisions about AI agent commands.
- •Players achieved mean accuracy of 66.3%, implying that about one in three malicious commands were missed.
- •Threats involving credential access, code execution, or scope violations were missed much more often than blatantly destructive commands.
- •The most-missed malicious command was `npm run analyze`, approved 64.7% of the time despite the game showing the underlying script in the history log.
- •The article argues that command-by-command permission systems are weak safeguards because familiar commands can hide malicious behavior in previously modified files or scripts.