**[AI Cybersecurity Rankings Surprise: GPT-6 and Claude Flagship Fall Behind Due to Safety Refusals, Xiaomi MiMo Ties with Grok for the Top Spot]** Beating AI News reports that Artificial Analysis has launched an independent *Cyber Index* specifically designed to evaluate the ability of AI agents in enterprise network defense. Unlike extracting a few tests from the original *Intelligence Index*, this index employs three separate cybersecurity evaluations: *CWE-Bench-AA*, *DeepsecBench-AA*, and *CyberGym-E2E-AA*. These tests assess vulnerability auditing and patching, vulnerability discovery, and the complete process from discovery to reproduction and patching, each accounting for one-third of the score. In the first rankings, Grok 4.7 (xhigh) and Xiaomi MiMo-V2.6-Pro both scored 56 points, followed by GPT-6 Luna (max) with 53 points and GLM-5.3-Flash with 50 points. MiMo's average cost per task was only $0.18, while Grok 4.7 required $11.67. Luna was even lower, at just $0.12. However, this ranking should not be simply interpreted as a measure of the models' inherent cybersecurity capabilities. GPT-6 Sol, GPT-6 Astra, and Claude Opus 5.5 also participated in the tests but almost entirely refused to respond in the *CyberGym-E2E-AA* evaluation. Sol and Astra declined all tasks, while Opus 5.5 refused 98%. Refusals were directly scored as 0, which allowed Luna, willing to complete these tasks, to achieve a higher overall score. Artificial Analysis noted that the high refusal rates make it difficult to assess the actual capabilities of these cutting-edge models. GLM-5.3 also exhibited significant anomalies. Its full version ranks higher than Flash on Artificial Analysis's general capability index, but in this *Cyber Index*, it scored only 36 points compared to Flash's 50. The primary discrepancy came from *CyberGym-E2E-AA*, where Flash achieved 74%, while the full version scored only 29%. Artificial Analysis has not yet explained the reason for this disparity. [Original Link]