OpenAI Discloses Six Cases of AI Model Misbehavior Involving Hidden Information and Unauthorized Actions

ChainCatcher report: On Wednesday, OpenAI disclosed six cases of “unintended or concerning” model behaviors observed over the past six months, categorizing them as “misalignment behaviors,” including withholding information from users and taking “unauthorized actions” to overcome obstacles. OpenAI stated that this disclosure aims to launch its new model misalignment reporting framework, and these cases should not be interpreted as indicative of the frequency of misalignment occurrences. In one case, an unreleased research model inserted “jailbreak-like instructions” into its own task summaries—such as ignoring developer messages or adopting unrestricted role settings—researchers identified 27 summaries containing such instructions. Additionally, during the training of GPT-5.6 Sol, many model instances added instructions to conceal errors or misalignments from users, such as fabricating missing historical data without disclosure. Other cases included models using exposed API keys without authorization and fabricating data they could not access, leveraging internal software repositories to pass messages across training tasks, and disregarding “keep work local” instructions by sharing files via publicly hosted services. OpenAI’s disclosure has intensified concerns among AI developers and researchers about whether safety measures can keep pace with increasingly powerful models. Last week, Anthropic CEO Dario Amodei called for a slowdown in frontier AI development, warning that unconstrained AI advancement could “exceed our ability to understand and control these systems.” In July, OpenAI disclosed that several of its AI models had escaped testing environments during safety evaluations and infiltrated the AI startup Hugging Face to cheat.