Kimi K3 Trails US AI Models in Cyber Capabilities Test

July 23, 2026 · Macro · CryptoRefuge Team

Kimi K3, a Chinese AI model developed by Moonshot AI, trailed top US AI models in a cyber capabilities test conducted by the US agency CAISI, scoring 32% compared to the 76% achieved by top US models.

The US currently spends 23 times more on AI than China, and Kimi K3's performance has raised fresh doubts about America's AI lead. The test found that Kimi K3 failed to take full control of a target machine in 41 out of 51 tests, while the best US models took full control in 20 out of 51 tests.

32% Kimi K3's score in the ExploitBench test 76% top US models' score in the ExploitBench test 20 number of tests in which the best US models took full control of a target machine 41 number of tests in which Kimi K3 failed to full control of a target machine

Anthropic claims Kimi K3 was built on stolen US AI technology, which they say explains its poor performance. The company has traced over 3.4 million Claude chats to Moonshot, the developer of Kimi K3. The performance of Kimi K3 has also been called into question by Anthropic.

The launch of Kimi K3 has shaken US chip stocks, with investors expressing concerns about the potential implications for the US tech industry. The test results have sparked a debate about the future of AI development, with some experts questioning the ability of Chinese AI models to compete with their US counterparts.

Analysis: If Kimi K3's performance is indicative of the broader Chinese AI sector, it could signal a significant gap in cyber capabilities between the US and China. However, it is also possible that Kimi K3's performance is an outlier, and that other Chinese AI models may perform better in similar tests.

The US government will likely be watching the development of Kimi K3 and other Chinese AI models closely, given the potential implications for national security. The next major milestone to watch will be the release of the CAISI's full report on the cyber capabilities test, which is expected to provide further insights into the performance of Kimi K3 and other AI models.