360’s AI Security System Tops CyberGym With 99.4% Vulnerability Reproduction Rate

The headline number matters less than the company it puts on the board: a Chinese security vendor claiming a top spot against agents from OpenAI, Anthropic and Google on a benchmark framed as AI security's "World Cup." That framing positions automated vulnerability reproduction as the competitive frontier, which shifts attention from model capability to who can operationalize offensive security tooling. The open questions are whether independent evaluations corroborate the 99.4% figure and whether CyberGym results translate into real-world disclosure and patching workflows.