US AI Security Report Claims Lead Over Kimi K3 Despite Unequal Testing Conditions

Commerce Dept report states US leads AI security, but Kimi K3 faced limited testing scope. US models scored 76% vs Kimi's 32% on benchmarks.

Woofun AI reports that the U.S. Department of Commerce’s AI Standards and Innovation Center, alongside the UK’s AI Security Institute, evaluated Kimi K3’s network attack capabilities. The assessment concluded that the United States maintains a lead, though the report acknowledged significant disparities in testing conditions. Kimi K3 participated in a restricted subset of tests due to hosting constraints, with its performance estimated primarily through 41 vulnerability exploitation benchmarks, whereas competing U.S. models underwent more comprehensive evaluations.

Consequently, the report notes a larger margin of error for Kimi K3’s results. In vulnerability exploitation, Kimi K3 achieved a 32% score, surpassing GLM-5.2’s 24% but falling short of the approximately 76% average for leading U.S. models. During simulated attack chain scenarios, Kimi K3 completed an average of 17 out of 32 steps, breaching a network once in ten attempts, compared to the 28.5 steps completed by top U.S. models. The evaluation identified autonomous attack capabilities in Kimi K3 and noted failures in its security safeguards, while reiterating the limited scope of the overall testing framework.