“The Chinese AI model KIMI K3 scored 32.2% on the Exploit Bench benchmark, compared to an average of 76.2% for frontier U.S. models.”