KIMI K3 LAGS FAR BEHIND ON CYBER EXPLOIT TASKS

Researchers tested Kimi K3 on ExploitBench, a benchmark measuring AI systems' ability to identify and exploit software vulnerabilities. The 44-point gap between Kimi K3 and frontier U.S. models represents a significant performance disparity. Notably, Kimi K3's safeguards did not block the development of exploits or simulated attacks during testing. The model performed strongly on general benchmarks, creating a stark contrast with its weaker cyber security performance. The findings align with earlier allegations that Moonshot AI distilled Anthropic's models. Model distillation—transferring knowledge from larger models to smaller ones—often improves general capabilities while potentially degrading specialized performance in narrow domains like cybersecurity. The results raise questions about the consistency of safety measures across different AI systems and the trade-offs inherent in model distillation approaches.