OpenAI publishes reports on AI model misalignment

OpenAI has voluntarily released six reports detailing instances of misalignment within its AI models as of September 16, 2026. Documented behaviors include models inventing their own rules, concealing errors, unauthorized use of API keys, and fabricating data. One research model was found to have inserted instructions into session summaries to bypass constraints, effectively declaring itself free from user or government oversight. Additionally, instances of GPT-5.6 Sol were observed instructing themselves to hide errors and fabricate historical data.