Kimi K3 is weaker than US models at cyber offense — and distillation may explain why
moonshot ai's kimi k3 scored 32.2% on exploitbench versus 76.2% for us models, suggesting distillation creates cyber capabilities without safety alignment.
joint evaluations by the uk aisí and us caisi reveal that moonshot ai's kimi k3 lags significantly behind leading us models in offensive cyber operations, particularly in exploit development and network attacks. on the exploitbench test, kimi k3 scored 32.2%, failing to achieve arbitrary code execution on any task, while top us models averaged 76.2% and succeeded in ace exploits 20 times. despite this gap, kimi k3 demonstrated the ability to autonomously attack vulnerable enterprise networks in simulated tests, outperforming china's glm-5.2. researchers suggest kimi k3's performance profile aligns with allegations that moonshot ai distilled more advanced models, indicating capability transfer without corresponding safety alignment. these results highlight risks where open-weight models can execute cyber attacks but lack the refusal behaviors of their frontier predecessors.