Kimi K3 vs. U.S. Models: 32% vs. 76% on Cyberattack Benchmarks

The UK AI Safety Institute (AISI) and the U.S. Center for AI Standards and Innovation (CAISI) jointly tested Moonshot AI's Kimi K3 on its ability to conduct cyberattacks. This is the first publicly released result in which two governments simultaneously measured the gap between Chinese and American models on practical offensive AI capabilities.

On ExploitBench — a benchmark covering 41 Chrome V8 vulnerabilities — Kimi K3 scored 32.2%. Leading U.S. systems scored 76.2%; GLM-5.2 (another Chinese model) scored 24.4%. The most telling metric was Arbitrary Code Execution: U.S. models achieved it on 20 of 41 tasks; Kimi K3 achieved it on none.

The second test — "The Last Ones" — simulated a real-world corporate attack across 32 sequential steps. Kimi K3 reached step 17 on average; U.S. systems reached step 28.5. Kimi K3 completed the full 32-step chain just once in ten attempts. Meanwhile, the model's safety filters blocked neither exploit development nor simulated offensive operations.

One of the report's key findings concerns the likely role of distillation. Kimi K3 was presumably trained on the outputs of other models rather than on real cyberattack examples. This would explain strong performance on general benchmarks alongside weak specialized capability: the model can reason, but it cannot attack in practice.

The results matter not as absolute scores but as a precedent: Washington now has published data that quantitatively confirms the effect of AI chip export controls. The 44-percentage-point gap between Kimi K3 and U.S. leaders is a measurable trace of years of sanctions policy.