How would you conduct adversarial testing to proactively uncover vulnerabilities and failure modes in our models at Character.AI?
