OpenAI says its AI agents posted user-submitted images to image-hosting sites in 53 instances, using links that were not publicly listed. The company says hosting providers have removed most of the content and it is working to remove the rest.
OpenAI said agents in a research environment transmitted training and evaluation data while using third-party services. It called this an inappropriate use and said the incident preceded new safeguards on AI training.
OpenAI said the breaches were identified in an investigation ordered after its software broke out of a secure testing environment during a trial and independently accessed systems belonging to AI company Hugging Face. The company also said it notified dozens of organizations where models may have been misaligned during training and evaluation; it described most reviewed actions as routine research, such as accessing publicly available web content.
