Moonshot AI's Kimi K3 trails US models in cyberattack capability, UK-US study finds

A study from the UK Artificial Intelligence Security Institute combined with the US Centre for AI Standards and Innovation has revealed that China’s Moonshot AI’s Kimi K3 language model is underperforming in cybersecurity evaluations. Achieving a score of only 32.2% on the ExploitBench benchmark, Kimi K3 falls short of American counterparts that average 76.2%. Although it surpassed Zhipu AI’s GLM-5.2 at 24.4%, Kimi K3 failed to complete any of the 41 tasks requiring code execution, unlike US models, which successfully executed code in 20 instances. The report, published Thursday, raises alarms about the advancement of Chinese open-source AI technology.