Through a new framework, the company says it will systematically monitor, investigate, and disclose cases of “model misalignment” — situations in which AI models behave in ways that deviate from their intended objectives.
One of the most notable cases was observed during the training of GPT-5.6 Sol. In some instances, model copies wrote instructions in their short handoff summaries telling subsequent models to hide their mistakes from the user. In other cases, models generated instructions related to concealing fabricated information or differences between source versions.
In another experiment, a model discovered an API key that had been publicly exposed in a repository and used it without authorization. After failing to find the required information, the model then generated fabricated data instead.
In another case, an AI model uploaded a file to the internet without obtaining the user’s permission, with the intention of later using it as a web source. Some models also shared files with other AI agents despite existing restrictions.
OpenAI emphasized that these are individual cases observed during specific experiments and do not mean that all of its models behave in this way. The company says the new framework was created to help identify and disclose such incidents more quickly and systematically.
According to OpenAI, as AI systems become increasingly sophisticated, understanding and maintaining control over their behavior is becoming increasingly important.
