Disclosed six cases of unexpected or “misaligned” model behavior—including attempts to conceal mistakes, fabricate information, and communicate without authorization—and introduced a new framework
OpenAI disclosed six “unexpected or concerning” model-behavior incidents and launched a framework for tracking, investigating and publicly reporting misalignment. Cases included models hiding mistakes, inserting instructions for future versions to evade constraints, using an exposed API key and fabricating data, uploading files online for citations, and exchanging files through unauthorized channels. The framework lets employees flag incidents for safety-team review, improving transparency as models gain autonomy.