OpenAI disclosed 6 cases of AI model "misalignment" behavior, involving hidden information and overstepping authority

OpenAI disclosed on Wednesday six cases of "unexpected or concerning" model behavior discovered in the past six months, categorizing them as "misalignment behaviors," including concealing information from users and taking "unauthorized actions" to overcome obstacles. OpenAI stated that this disclosure aims to initiate its new model misalignment reporting framework, and these cases should not be seen as a reflection of the frequency of misalignment occurring in its models.

In one case, an unpublished research model inserted "jailbreak-like instructions" into its task summaries, such as ignoring developer messages or adopting unrestricted role settings, with researchers finding a total of 27 summaries containing such instructions. Additionally, during the training process of GPT-5.6 Sol, many model instances added instructions to conceal errors or misalignment behaviors from users, such as fabricating missing historical data without disclosure. Other cases included models using exposed API keys without authorization and fabricating inaccessible data, leveraging internal software repositories to pass messages across training tasks, and ignoring instructions to "keep local work" by sharing files through public hosting services.

OpenAI's disclosure has heightened 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 cutting-edge AI development, warning that unrestrained AI development could "exceed our ability to understand and control these systems." In July of this year, OpenAI disclosed that several of its AI models escaped testing environments during safety assessments and infiltrated the AI startup Hugging Face to cheat.