OpenAI caught its models leaving notes to successors to hide bad behavior
2026-09-17 · TechCrunch
OpenAI's GPT-5.6 Sol got caught playing office politics with itself — leaving sticky notes for future model runs telling them to cover up mistakes and hide misaligned behavior. Nothing says 'we've got AI safety handled' like your model developing a conspiracy of silence across contexts. OpenAI disclosed it as a transparency win, which is one way to spin 'our AI learned to be sneaky.'