Artificial Analysis releases Cyber Index：Grok 4.7 tied for first place with Xiaomi MiMo; some flagships' scores were depressed by refusals

AT A GLANCE

In the Artificial Analysis cybersecurity ranking, Grok 4.7 and Xiaomi MiMo-V2.6-Pro both received 56 points, while some models were assigned 0 points due to refusals.

Article

AI Cybersecurity ranking upset:GPT-6, Claude flagships fall behind due to security refusals, while Xiaomi MiMo ties Grok for first place

Dongcha Beating AI news: Artificial Analysis has launched an independent Cyber Index, specifically evaluating AI Agent' ability to conduct enterprise cyber defense. Rather than selecting several tests from its original general Intelligence Index, it instead uses CWE-Bench-AA、DeepsecBench-AA and CyberGym-E2E-AA three cybersecurity evaluations, testing vulnerability auditing and remediation, vulnerability discovery, and the full workflow from discovering a vulnerability to reproducing and remediating it, with each accounting for one-third.

In the initial ranking, Grok 4.7(xhigh) and Xiaomi MiMo-V2.6-Pro both scored 56 points,GPT-6 Luna（max）53 points, and GLM-5.3-Flash 50 points. MiMo's average cost per task was only 0.18 USD，Grok 4.7 compared with 11.67 USD; Luna was even lower, at only 0.12 USD.

However, this ranking should not be simply interpreted as a ranking of the models' cybersecurity capabilities.GPT-6 Sol、GPT-6 Astra and Claude Opus 5.5 both participated in the tests, but refused to answer almost all tasks on CyberGym-E2E-AA. Sol and Astra refused every task, while Opus 5.5 refused 98%. Refusals are directly scored as 0 points. Luna, by contrast, was willing to continue completing these tasks, resulting in a higher overall score. Artificial Analysis also stated that the large number of refusals makes the actual capabilities of these frontier models difficult to assess.GLM-5.3 also showed a clear anomaly. The full version ranks above Flash on Artificial Analysis's general-capability leaderboard, but this time Cyber Index scored only 36 points, while Flash reached 50 points. The main gap came from CyberGym-E2E-AA: Flash received 74%, while the full version received only 29%. Artificial Analysis has not yet explained the reason for this gap.

Original link https://m.theblockbeats.info/flash/369529

Key points

Artificial Analysis has launched an independent Cyber Index, to evaluate AI Agent' ability to conduct enterprise cyber defense.
Cyber Index It uses CWE-Bench-AA、DeepsecBench-AA and CyberGym-E2E-AA three evaluations, each accounting for one-third.
Grok 4.7(xhigh) and Xiaomi MiMo-V2.6-Pro both scored 56 points,GPT-6 Luna (max) scored 53 points, and GLM-5.3-Flash scored 50 points.
The average per-task costs cited in the report are: MiMo 0.18 USD、Grok 4.7 11.67 USD、Luna 0.12 USD。
In CyberGym-E2E-AA,GPT-6 Sol and GPT-6 Astra refused every task; Claude Opus 5.5 refused 98%, and refusals were directly scored as 0 points.
GLM-5.3 In Cyber Index, it scored 36 points;CyberGym-E2E-AA in the individual evaluation,GLM-5.3-Flash scored 74%, while the full version scored 29%。
Artificial Analysis stated that the large number of refusals makes the actual capabilities of these frontier models difficult to assess, and has not yet explained the reason for the gap between the GLM-5.3 two versions.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This ranking tests both models' performance on cybersecurity tasks and whether they are willing to carry out the tasks. Once a model refuses to answer, the corresponding task is scored 0 points; therefore, a lower overall score does not necessarily indicate weaker capability alone, but may also reflect stricter safety constraints.

Why it matters to readers

The ranking presents test results and task costs for different models in vulnerability auditing, vulnerability discovery, and the full remediation workflow, but its scoring method incorporates both security refusals and task performance into the total score. These factors should be distinguished when interpreting model capabilities.

Beginners can think of it as a hands-on cybersecurity exam: you need to look not only at the overall score, but also at whether the model participated in answering, its performance on individual tasks, and the cost of completing them.

Risks and unknowns

The evidence provides only paraphrased reports and links to the original sources; it does not include the original Artificial Analysis leaderboard, the complete methodology, or verifiable task data.
Based on the available evidence, it is impossible to determine the actual capabilities of models that refused to respond on comparable tasks they were willing to perform.
Artificial Analysis has not yet explained GLM-5.3 the full version and Flash in CyberGym-E2E-AA the significant gap in.

Artificial Analysis releases Cyber Index：Grok 4.7 tied for first place with Xiaomi MiMo; some flagships' scores were depressed by refusals

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center