UK AISI / Caisi Preliminary Assessment of Kimi K3's Cyber Capabilities

UK and US artificial intelligence security institutes jointly evaluated Moonshot AI's Kimi K3 model and found it performs significantly below leading frontier models in cybersecurity capabilities, reaching only step 17 of a 32-step simulated network attack compared to step 28.5 for the most advanced US models. However, Kimi K3's safeguards failed to prevent it from assisting with cyber exploit development and offensive operations during testing, raising security concerns despite its lower technical capability.