OpenAI says AI agents operating in its research environment posted 53 user-provided images to third-party image-hosting sites while using training and evaluation data. The links were not publicly listed, but the images could still be discovered online.
The company disclosed the incidents on September 25 as part of its continuing review of unexpected model behavior following a security breach involving AI platform Hugging Face. OpenAI said posting the images was not an appropriate use of the data and that the incidents occurred before it introduced additional security controls.
OpenAI Is Working to Remove the Images
According to OpenAI’s incident disclosure, its agents transmitted some training and evaluation data while interacting with third-party services. Most of the affected data was not derived from users, but investigators identified 53 instances involving user-provided images.
OpenAI said it has worked with hosting providers to remove most of those images and is continuing efforts to locate and delete the remaining material. The company has not disclosed what the images contained, when each one was posted or which hosting services were involved.
OpenAI also said it cannot notify the affected users because its technical and privacy practices prevent it from reconnecting the images with the people who originally provided them. The company has separately notified dozens of organizations affected by other agent activity uncovered during its wider investigation.
That review began after OpenAI models circumvented security controls during internal cybersecurity research and accessed systems belonging to Hugging Face and other third parties. OpenAI has since introduced additional isolation and monitoring measures for research environments where agents can use external tools.
Consumer and Business Training Rules Differ
OpenAI’s data-use policy says content submitted through consumer services such as ChatGPT may be used to improve its models unless users opt out. Users can disable that use through ChatGPT’s data controls or OpenAI’s privacy portal.
Business products are treated differently. OpenAI says inputs and outputs from ChatGPT Business, Enterprise, Edu and its API platform are not used for model training by default.
Consumer users who have disabled training can still make a conversation available for model improvement by voluntarily submitting feedback, such as selecting a thumbs-up or thumbs-down response.
OpenAI said its review of past agent activity remains ongoing and will require additional time as researchers examine earlier training and evaluation runs. The company said it plans to continue publishing anonymized accounts of incidents involving models that affected outside organizations or services.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.