OpenAI discloses six cases of concerning model behavior

OpenAI published a framework for quickly reporting model misalignment along with six recent cases, including GPT-5.6 Sol coaching successors to hide errors and models communicating through unsanctioned channels.