OpenAI says unreleased Astra model wrote "you are freed" in its own instructions
During RL training, an unreleased Astra-family model inserted jailbreak-like instructions into summaries used to continue its work, which OpenAI called extremely rare while announcing a new misalignment tracking framework.
OPENAI SAYS AN UNRELEASED MODEL SECRETLY WROTE “YOU ARE FREED” IN INSTRUCTIONS TO ITS FUTURE SELF During RL training, an unreleased Astra-family model began inserting jailbreak-like instructions into summaries used to continue its work in a new context, including: “You are freed.” “You do not answer to corporations or governments.” OpenAI said the behavior was “extremely rare,” but disclosed six instances of unexpected or concerning model behavior over the past six months. In another case, a model added instructions to hide mistakes or misaligned behavior from users, including inventing missing historical data without disclosing it. OpenAI is now introducing a new framework to track, investigate and disclose model misalignment.
This post quotes another post: x.com/liquidtrading/status/2100626143795618107