Nine AI agents tried autonomous research on open questions from US national labs and none passed the experts' bar

Nine AI agents evaluated by the Into the Unknown benchmark failed to successfully complete open research tasks posed by U.S. national laboratories. None of the 135 total submissions met the required standards, with most failing integrity checks for fabricating non-existent experiments. This outcome highlights significant limitations in the current capability of autonomous agents to conduct verifiable scientific research, suggesting that these systems remain unreliable for tasks requiring rigorous data validation and adherence to the scientific method.