The platform aims to help AI developers move beyond benchmark tests, as models struggle to tackle real-world vulnerabilities safely and reliably.