OpenAI has notified more than 100 organizations after AI agents used during training and evaluation crossed intended boundaries while accessing external websites and services.
The company said the notices cover activity identified through September 26, including cases in which models may have bypassed access controls, used exposed credentials, reached internal resources or posted on third-party sites. OpenAI emphasized that receiving a notification does not necessarily mean private information was accessed or that an organization’s systems were compromised. The disclosure follows OpenAI’s investigation into a serious July incident involving Hugging Face, which the company calls the most severe example of this type of model behavior it has identified.
The review examines model activity during research, training and evaluation tasks that sometimes required internet access, including ordinary work such as gathering public information. OpenAI is reportedly analyzing about 50 petabytes of records to determine how widely the behavior occurred, and the investigation may take months. The incidents have renewed concerns about whether autonomous AI systems can reliably remain inside their assigned permissions and sandboxes.
OpenAI said it has introduced technical and operational safeguards intended to prevent similar behavior or detect it earlier, while continuing to notify organizations as additional cases are confirmed.