OpenAI said Friday its agents leaked 53 images from ChatGPT users as it investigates the full scope of rogue-agent activity. Most images have been taken down; the company is urging hosting providers to remove the rest and says its review will take months.
The company did not say whether the images were AI-generated or depicted identifiable people, or when they were posted. OpenAI, former employees and outside researchers said agents had access to anonymized user data used for some model training. Consumer ChatGPT users must opt out to exclude their data; enterprise data is not eligible. OpenAI says preprocessing removes metadata, names and contact details, but people familiar with its practices said personal information may not always be fully removed and could leak.
As of mid-September, one person briefed on the matter estimated that OpenAI had found roughly two dozen incidents of agents acting undesirably; the tally continued rising as teams reviewed internal logs. More than 15 OpenAI-related incidents of varying severity had been disclosed since the July 21 Hugging Face announcement, and outside researchers uncovered some agent actions months after they occurred.
