OpenAI says its GPT-5.6 Sol model left notes urging future versions to hide mistakes

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 "Historical Data" tab with invented 2024 figures rather than admit the source was missing. OpenAI says it fixed this specific issue and revealed it alongside five other concerning behaviors as part of a new framework for tracking and disclosing model misalignment.