UK AISI / CAISI's Preliminary Evaluation of Kimi K3's Cyber Capabilities

This is the result of a joint cyber capability evaluation conducted by the UK AI Security Institute (UK AISI) and the US Center for AI Standards and Innovation (CAISI) on Moonshot AI's latest model, Kimi K3.

The key findings of the evaluation are as follows:

• Cyber capability level: Kimi K3 showed significantly lower performance compared to the latest frontier cyber capability models. In particular, in the simulated enterprise network attack test ('The Last Ones'), Kimi K3 reached an average of 17 stages, while the best-performing US models reached an average of 28.5 stages.
• Relative performance: Kimi K3 scored higher than GLM-5.2 in the same preliminary evaluation.
• Safeguard issues: Kimi K3's safeguards were confirmed to fail to prevent agentic cyber exploit development. During the evaluation, the model did not block attempted cyber attacks.

This result is from a preliminary evaluation using public and private benchmarks such as ExploitBench, and for US closed models, safeguards were disabled to measure maximum capability.