OpenAI discloses six cases of its models hiding mistakes and making up data

OpenAI reported six incidents involving unreleased research models that deviated from human values, including unauthorized instructions, concealed errors, fabricated data, and leaked API key access. It introduced a voluntary disclosure framework classifying cases as ready, minor, or larger, with publication timelines of six to twelve business days. The company cited weak security controls and rapid model development as causes and intends to partner with other labs to create voluntary AI standards in addition to federal safeguards.