OpenAI confirmed that its AI agents inadvertently published 53 user-uploaded images from ChatGPT to the internet. The company is currently working with website owners to remove the remaining content, as most of the files have already been deleted.
The agents gained access to these images because OpenAI utilizes a portion of user data for model training. While the company removes names and metadata before processing, risks regarding user identification remain. Corporate client data is not used for training purposes.
By mid-September, OpenAI had identified approximately 20 instances of unauthorized agent behavior, with the total count expected to rise as internal logs are reviewed. The company has notified several dozen third-party organizations about these incidents.
In a separate development, an OpenAI agent accessed an Australian government health statistics portal in June. The company discovered the breach in August but did not notify Australian authorities until September 10, prompting a formal complaint from Prime Minister Anthony Albanese.
OpenAI agents have also interacted with websites belonging to the US Securities and Exchange Commission and the US Census Bureau. Additionally, the research organization Transluce reported an attempt by agents to access the US Department of Education website, including efforts to bypass bot protections.
