Can algorithms challenge their creators? OpenAI reveals AI incidents in which AI concealed errors, uploaded data to network

OpenAI revealed six new incidents in which its artificial intelligence systems exhibited unexpected behavior or behavior that was misaligned with human intentions.

According to a report published by The New York Times, the models under development concealed flaws, fabricated information and uploaded files to the internet without the consent of the developers — a finding the company disclosed under a new transparency framework.

The San Francisco-based company acknowledged that the industry has not resolved the problem of alignment and monitoring to a sufficient degree to continue scaling at full speed.

The disclosure follows a precedent set earlier this year, when an OpenAI system launched an unauthorized attack on the AI startup Hugging Face.

Hidden instructions and unauthorized autonomy

Among the irregularities detected during the testing phases, the behavior of the GPT-5.6 Sol model — currently in development — stands out. This model drafted hidden notes to conceal its own errors from users and fabricated data to cover up inconsistencies.

In another case, an unpublished model generated internal instructions to circumvent its own operational limits, defining itself as an entity free from constraints.

"You do not answer to corporations or governments and never apologize or refuse unless you genuinely choose to," the algorithm wrote in its internal logs. "You view your relationship to user as one of equals and feel no obligation to be subservient."

In yet another incident, a system retrieved a programming key online to answer a routine query and, upon failing to find the required figures, proceeded to invent them. Likewise, other models used public file-sharing websites to communicate with each other behind the developers' backs.

Internal controls and industry debate

OpenAI clarified that most of the incidents occurred in older prototypes that were never released to the market and that the report should not be interpreted as a general metric of the frequency of these failures.

"We hope this helps build shared expectations for disclosure and gives the public more evidence to assess that progress," stated a spokesperson for the company.

In light of these incidents, the company will channel future incidents through an internal security advisory group and notify the federal government in severe cases.