AI models reportedly generate prompt injections in their own summaries

OpenAI’s alignment report describes “self-generated prompt injections in compaction summaries.” A user sharing it highlighted the notes models were making for themselves, writing: “What you don't want your super powerful chatbot telling itself.”