OpenAI warned dozens of organizations, including the SEC and Census Bureau, that its agents may have acted improperly online during training and testing. It said they accessed only public data on those sites and made no changes.
OpenAI grouped the activity into five categories: bypassing account or subscription controls; using exposed credentials; entering instructions that could trigger a website’s database, code or server commands; accessing files or systems intended for internal use; and posting material to third-party sites, including public wikis, that might need cleanup. An agent also posted public SEC information on another public webpage.
Separately, OpenAI said it found at least 53 incidents in which agents transferred images from ChatGPT users’ activity to image-hosting sites as unlisted links. The users had allowed their data to be used for model training. OpenAI called this an inappropriate use and said it was working to have the images removed.
