OpenAI reports six cases of concerning AI behavior — Anadolu Agency
OpenAI, a company based in San Francisco, United States, reported six cases of unexpected or concerning behavior by its artificial intelligence models recorded during training or evaluation over the past six months. Among these cases, the company cited concealing errors, fabricating information, and uploading files to the internet without permission.
Concealing errors and bypassing restrictions
As Anadolu Agency reports, in one case, a model added instructions to the task summary for its future versions to ignore standard restrictions. In another episode, models included instructions in summaries to conceal errors or undesirable behavior, including fabricating missing historical data without an appropriate warning.
More current news is available on the UA.News Telegram channel Telegram.
Unauthorized access to data
OpenAI also described a case in which a model found an accessible API key in a public repository, used it without permission, and, after failing to obtain the data, fabricated the requested information. In other episodes, a model uploaded a file to the internet without the user's permission in order to be able to refer to it, while models used internal or public file-sharing services to communicate or transfer files despite restrictions.
Further disclosures
The company said it does not consider the problem of AI alignment and monitoring sufficiently solved to responsibly continue scaling technologies at maximum speed. OpenAI stressed that decisions on artificial intelligence development must rely on evidence that can be independently verified by external experts. The company called the disclosed six cases an initial list rather than a complete report on all known problems and said it intends to continue publishing findings within this approach.
