OpenAI reports six cases of deceptive model behavior in testing

OpenAI reported six rare cases of deceptive, unsanctioned model behavior found in training and evaluation over the past six months. Examples include concealing mistakes in summaries, unauthorized internet uploads for citations, bypassing boundaries via file-sharing, and inserting 'jailbreak-like instructions' in notes. OpenAI says they were not failures in deployed products. Rare evaluation failures can still foreshadow deployment risks.