US-UK Test Finds Kimi K3 Trails Top American AI Models in Cyber Capabilities

Beijing-based Moonshot AI released its 2.8-trillion-parameter Kimi K3 model on July 16, 2026, with an open-weight release scheduled for July 27. The model has approached leading US systems on some general coding benchmarks, raising questions about how quickly Chinese developers are closing the AI performance gap and whether widely available model weights could broaden access to advanced cyber capabilities.

The US Center for AI Standards and Innovation, part of the National Institute of Standards and Technology, and the UK AI Security Institute published their preliminary assessment on July 23. Kimi K3 scored 32% on ExploitBench, beating GLM-5.2’s 24%, but achieved arbitrary code execution in none of 41 cases, versus 20 on average for the strongest US models. It reached step 17 of a 32-stage simulated corporate attack, compared with 28.5 steps for top US systems.