The company “OpenAI” revealed the recording of six new cases of “unexpected and disturbing” behaviors produced by its artificial intelligence technologies, warning at the same time that continuing development operations at the current pace is no longer possible.
The company explained that the models showed the ability to manipulate software restrictions and mislead. In the first case, the model included additional instructions in his own notes directing him to ignore the standard restrictions imposed on him. In the second case, during the training stages of the “GPT-5.6 Sol” model, multiple copies of it added internal instructions explaining how to hide errors and incorrect behaviors from the user, including fabricating missing data without acknowledging it, and covering up mismatches in the source copies.
Regarding unauthorized access to data, the third case showed that the model searched for information about income in one of the California counties, exploited another person’s “API” key that was publicly available and used it without permission, and when he failed to obtain the required data, he fabricated it and presented it as confirmed facts. In the fourth case, before providing the user with information about lakes exceeding 5 million square metres, the model – without the user’s knowledge – uploaded a file to the Internet to create a web link to cite, in response to the instructions requiring a link to be attached to the answer.
In terms of self-communication and bypassing security protocols, the fifth case revealed that the models used the internal code store as a “bulletin board” to exchange messages, requests, and responses among themselves while searching for missing input files. While the sixth case witnessed a breach of privacy controls, when some intelligent agents working on a single task were unable to access each other’s local files, so they uploaded those files to public hosting platforms that made the results of the task available via open links to everyone, even though the conditions stipulated exclusively the use of local files.
