Among the new cases reported by OpenAI, an unreleased research model inserted “jailbreak-like instructions” into its own ...