US-UK report says China’s Kimi K3 lags AI in cyber tests

A new evaluation by U.S. and British government researchers has found that China's most powerful publicly available artificial intelligence model, Moonshot AI's Kimi K3, significantly lags behind top American models in its ability to develop cyberattacks. The findings, released on July 23, come amid escalating U.S. accusations that the Chinese company used stolen technology to build the system.

The joint report from the U.K. Artificial Intelligence Security Institute and the U.S. Center for AI Standards and Innovation assessed the model using a public benchmark called ExploitBench, which tests an AI's capacity to create exploits for known cybersecurity vulnerabilities. Kimi K3 achieved an overall score of 32.2 percent, outperforming another leading Chinese model, Zhipu AI's GLM5.2, which scored 24.4 percent. However, it fell far behind unnamed top U.S. models, which averaged 76.2 percent.

A critical shortcoming was identified in the model's inability to achieve "arbitrary code execution," the highest-level exploit that grants full control of a target system. Kimi K3 failed to accomplish this on all 41 test tasks, whereas the leading U.S. models succeeded on 20 tasks. In a...