Hacker News new | past | comments | ask | show | jobs | submit
thayne | on: Humans missed 1 in 3 threats approving AI agent commands across 40k game runs
Also a lot of them may or may not be safe depending on additional context that you don't have in the test.