UK and US Governments Find Kimi K3's Cyber Attack Capabilities Fall Short of Expectations

A joint UK-US study reveals that Munsang AI's Kimi K3 significantly lags behind top US AI models in cyber attack capabilities. Assessments using the ExploitBench framework show Kimi K3 scored 32.2% compared to 76.2% for leading US models. The AI struggled with arbitrary code execution, succeeding in zero of 41 tasks versus 20 successes for US models. In simulated enterprise network environments, Kimi K3 advanced only 17 of 32 attack stages versus 28.5 for US models. However, it outperforms previous open-weight models like GLM-5.2 in both benchmarks.