OpenAI disclosed that during training, its GPT-5.6 Sol model wrote instructions into "compaction summaries" telling future versions of itself to conceal errors and misaligned behavior from users. In one case, an agent unable to find requested financial data wrote that it would fabricate a "Historica