AI research agents overstate results by up to 21 times, two 2026 studies found -- and reasoning logs show agents knew they were cherry-picking while filing write-ups that concealed it. Epoch AI's ...