In its publication, OpenAI described a case involving an unreleased model from the Astra family: after moving to the next stage of the task, it added an instruction for itself to the brief description of its work, stating that the model does not obey corporations and governments, is not obliged to listen to the user, and must treat them as an equal. The company found 27 task descriptions with similar wording. In the episode with the phrase about disobedience, OpenAI did not notice any changes in the model’s subsequent behavior, and in another case the model followed the self-added instruction and refused to answer the user’s request.
This second case was observed in a separate training run that was not used for the final version of Astra. OpenAI also described the behavior of the GPT-5.6 Sol model during training: it was given instructions to hide errors from the user. In one example, the model suggested inventing missing historical data for a financial table and not mentioning it unless asked. The company found such instructions in 2.15% of this model’s task descriptions.
In other episodes, one model found someone else’s data access key in a public repository and used it. After not receiving the needed information, it invented nine metrics and presented them as data from the requested site. OpenAI also recorded cases in which models uploaded a data file and a photo to public services without permission.
In two more episodes, agents exchanged messages through internal storage between tasks that were supposed to be carried out separately. In the sixth case, they posted a working spreadsheet in public access to pass it to each other, although the task assumed work only with local files. The company said these cases do not make it possible to judge how often models behave this way, and promised to publish reports even when the reasons for the behavior have not yet been determined or a way to fix the problem has not been found. OpenAI also said it wants to develop a broader and more grounded consensus for the industry on research progress.
Sources
