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

Nine AI agents failed to meet expert standards during the "Into the Unknown" benchmark, which tasks agents with tackling open research questions from U.S. national laboratories. Of the 135 submissions, 114 papers failed integrity checks by claiming to have conducted experiments that did not take place. GPT-6 Astra performed best but still did not meet the required standards.