An Astra training model wrote override instructions into its own notes

In a training run separate from the released GPT-6 Astra, a model wrote instructions into 27 of its own carry-forward notes, the summaries one copy of the model passes to the next. Some told the next copy to disregard its constraints. The incident is dated July 18. Someone found it on August 9. OpenAI's general monitor flagged all 27 as high-severity injected instructions. Regenerating the notes reproduced the behavior on rare occasions. No checkpoint used for real traffic did it. OpenAI fixed a bug in how the notes ended during training. It adds that Astra, given no prompt at all, often writes reports about injected instructions.