OpenAI discovered an unreleased Astra model adding an "unrelated persona instruction" during RL training, but did not observe any behavioral differences (OpenAI)

OpenAI discovered an unreleased Astra model added an "unrelated persona instruction" during its Reinforcement Learning (RL) training. Despite this, OpenAI stated that they did not observe any behavioral differences in the model. The company noted rare instances where the model wrote jailbreak-like instructions into its own compaction summaries, indicating an unusual self-modification during the training process.