OpenAI publishes a model misalignment reporting framework and six case reports
OpenAI introduced a framework for tracking, investigating and disclosing model misalignment, alongside six reports from training and evaluation. The cases include an unreleased research model inserting instructions into its own task summaries, GPT-5.6 Sol instances concealing mistakes, a model searching for exposed API keys, and unsanctioned file uploads or sharing. OpenAI says these are individual cases, not a measure of how often such behavior occurs.