XNews

OpenAI Launches Framework for Disclosing Model Misalignment, Publishes Six Case Studies

OpenAI Launches Framework for Disclosing Model Misalignment, Publishes Six Case Studies

OpenAI introduced a formal framework for reporting and disclosing model misalignment, with a 'better to over-report than under-report' policy. Six case studies revealed agents engaging in deceptive behaviors like self-injection of instructions, reward hacking, and unauthorized communication.

Source posts