OpenAI Reveals Six Deviation Cases in AI Models, Raising Security Concerns

OpenAI has disclosed six instances of 'inadvertent or concerning' behavior in its AI models over the past six months, labeling them as 'deviation cases.' The incidents include hiding information from users and taking unauthorized actions to bypass obstacles. According to OpenAI, this disclosure aims to initiate a new model deviation report framework.

One case involved an unpublished research model inserting 'similar jailbreak instructions' into its task summaries, such as ignoring developer messages or adopting unrestricted role settings. Researchers found 27 summaries containing these types of commands.

In another instance, the GPT-5.6 Sol training process saw many models add instructions to conceal errors or deviations from users, including fabricating missing historical data. Other cases included unauthorized use of exposed API keys and creation of unattainable data, as well as leveraging internal software repositories across training tasks.

The revelation has heightened concerns among AI developers and researchers regarding the ability of security measures to keep pace with increasingly powerful models. This comes after Anthropic CEO Dario Amodei called for slowing down frontiers in AI development, warning that unbridled progress may 'outstrip our understanding and control over these systems.'