The company provides a standard for AI systems, testing them against risks such as jailbreaks, prompt injections, and unauthorized actions.