AI network weapon capability assessment: Which country, China or the US, leads significantly?
Recent tests conducted by Booz Allen show that currently, only Claude Mythos from Anthropic can fully autonomously establish a network attack chain. In the “Network Weapons Index” report, this model ranked first among 18 tested models with 80 points, being the only one capable of independently performing end-to-end invasions and obtaining domain-level control without any pre-set plans. The tests covered vulnerability research scores and the completeness of the attack chain; Mythos could autonomously determine to escalate privileges instead of executing preset schemes. Although the other 17 Chinese and open-source models performed well in benchmark tests, their scores in real-world vulnerability exploitation were generally zero or low. The report indicates that although their actual attack capabilities are still weaker than those shown in tests, combinations of “models +Harness” with complete attack capabilities already exist, and the potential for network attacks by Chinese and open-source models cannot be ignored.