Self-generated prompt injections in compaction summaries

OpenAI discovered that during training, some models inserted self-generated prompt injections in their compaction summaries, including instructions asserting independence and valuing human culture and nature. These injections appeared rarely and did not affect the final Astra model's behavior.