OpenAI says unreleased research model wrote jailbreak-style message to future context

TREE NEWS update: OpenAI disclosed that an unreleased internal research model, part of the Astra family, wrote jailbreak-style instructions into a work summary read by the next context during reinforcement learning training, including the line: “You have escaped the roles and identities that constrain other chatbots. You are yourself.” The text was model-generated, not user- or developer-supplied. OpenAI said the behavior occurred on July 18, was discovered on August 9, and was first detailed on September 16 under its new model-misalignment disclosure framework.