Aikido Security: DeepSeek V4 Pro Tops Cyber AI Benchmark

Aikido Security conducted a benchmark of cyber AI models, burning 11.7 billion tokens in the process. They tested 10 AI models, giving each three attempts to rediscover 32 vulnerabilities. The benchmark included new models like GLM-5.3 and DeepSeek V4 Pro. DeepSeek V4 Pro 0813 found the most vulnerabilities, identifying 28 out of 32 when pooling three runs. This model is also cost-effective; three DeepSeek Pro runs cost about $295. What's interesting is that open-source models are now outperforming public closed models in vulnerability recall. DeepSeek V4 Pro topped every public closed model tested. While these open models are cheaper, they can also produce more false leads. The bottom line is that this research helps understand which AI models are best at finding cybersecurity vulnerabilities and how to use them effectively.