OpenAI has confirmed that it is aware of a new security incident in which its AI agents uploaded user-provided images to third-party image hosting services.
According to OpenAI, most users were not affected as only 53 incidents were identified where agents accidentally uploaded images to the internet.
The disclosure is the result of a broader investigation by OpenAI into erroneous agent behavior following the Hugging Face security incident.
“As part of our ongoing investigation, we have identified instances where agents in our research environment submitted training and assessment data while using third-party services,” OpenAI noted in a Blog post.
“This does not constitute an appropriate use of this data and these cases occurred before we implemented the security measures described in our technical report.”
OpenAI says the vast majority of affected training and evaluation data did not come from users, but it found 53 cases involving user-provided images.
“While the vast majority of affected training and assessment data does not come from users, we have identified 53 cases to date where user-provided images were posted on image hosting sites as links that were not publicly listed,” OpenAI explained.
“We have successfully worked with the hosting providers to remove most of this content and are continuing to work to remove the remainder.”
According to OpenAI, data excluded from training by users or administrators was not affected
Some OpenAI training data may contain content from users who have consented to their interactions being used for training. However, the company says users who opted out were not affected.
“Any data that is not suitable for training because it is controlled by users or company administrators is not included,” OpenAI said. “For clarity, corporate or business account data and API usage are excluded unless an administrator has enabled them.”
OpenAI also states that additional steps are required before adding appropriate user data to training datasets.
“Before including eligible data, we take steps to protect privacy by separating it from account information and using a version of the OpenAI privacy filter to redact personal information such as names, contact information and account numbers.”
Following the incident, OpenAI said it has strengthened its training and assessment systems to make it more difficult for models to leak data via external services.
“As part of our response to our ongoing investigation, we have enhanced our training and assessment processes, including building security cases, securing and red-teaming our systems to prevent the model from filtering out data, and implementing additional monitoring,” the company noted.
The company continues to review senior agent activity month after month, starting with the Hugging Face incident, so more cases could emerge.
