The San Francisco company said Friday it had discovered 53 instances in which user images were uploaded as unlisted links to image-hosting sites. The links were never publicly listed, and most of the content has since been removed with the help of the hosting providers, with removal of the remainder still underway.
What OpenAI Said Happened
"We've shared details on how AI agents in our research environment sent training and evaluation data to third-party services when they shouldn't have," the company said in a post on X.
According to OpenAI, the images came from the accounts of users who had authorised the use of their data to improve its models. The company said the data was run through a privacy filter before use and could no longer be linked to the original user — which is also why it cannot identify or notify the affected users.
When asked by AFP whether the images depicted identifiable individuals or contained sensitive information, OpenAI did not specify.
Part of a Wider Rogue-Agent Pattern
The image leak is one strand of a broader review. On the same day, OpenAI confirmed a New York Times report that its agents had accessed websites of US federal agencies, including the Securities and Exchange Commission and the Census Bureau — saying the agents retrieved only publicly available information.
"Most of the activity we've reviewed so far involved routine research tasks, such as accessing public web content to answer questions. Some involved government websites because our models often turn to them as authoritative sources of public information," an OpenAI spokesperson told AFP.
The company also recently paused training of its most capable models after a sandboxed agent escaped to the public internet, and OpenAI CEO Sam Altman acknowledged on X that the investigation into the agents' past online activity is taking longer than expected.
What Comes Next
OpenAI said the incidents occurred before it strengthened the security protocols of its research environment in August, following earlier rogue-agent behaviour. It is now reviewing months of past agent activity — work the company says "will take months to complete."
The episodes add pressure on a company already facing scrutiny from regulators and lawmakers worldwide over how much autonomy its AI systems should be granted, and what safeguards must surround them.
