Kimi K3 trails US frontier models on cyber exploits; distillation may explain the gap

Kimi K3 在网络安全漏洞利用测试中大幅落后美国前沿模型，知识蒸馏或为原因

UK AISI and US CAISI tested Moonshot AI's Kimi K3 on offensive cyber tasks. It scored 32.2% on ExploitBench vs. 76.2% for leading US models and never achieved arbitrary code execution. On the TLO simulated network attack, K3 averaged 17 of 32 steps (US models: 28.5) and completed the full path only once in ten runs. Its safeguards did not block offensive use. The report notes the results are consistent with allegations that Moonshot AI distilled more advanced models.

Why it matters: Kimi K3's 32.2% on ExploitBench vs. the 76.2% US model average is a concrete, newsworthy gap, and the distillation hypothesis gives it a technical angle worth discussing. Score capped at 78 because it's a single-source report from the-decoder and key details sit behind a paywall.