53 user photos were exposed to external websites by rogue AI agents, according to OpenAI


OpenAI’s AI agents have once again come under scrutiny after the company confirmed that they accidentally exposed 53 images submitted by users on external image-hosting websites. Following the July Hugging Face incident and recent reports involving AI agents accessing Australian health websites, concerns over rogue AI systems have continued to grow. This latest case has raised particular concern because some of the images may have been personal photos that users never expected to appear outside their private conversations.

In a blog post, OpenAI said it had “identified 53 instances to date where user-provided images were posted” through links that were not intended to be publicly discoverable. In other words, the links could not be easily found through public internet searches, but anyone who obtained a link could potentially view the associated image.

OpenAI has not revealed whether the images were AI-generated or contained real people, nor has it disclosed when the images were first made available online. The company said most of the affected images have already been removed and that it is working with hosting providers to take down the remaining ones.

How did this happen

OpenAI uses anonymised user data to help train its AI models. Such data is meant to be processed to remove names, metadata and other information that could identify individual users. However, the process is not completely foolproof.

Researchers speaking to Reuters warned that sensitive details could occasionally remain in data, while AI agents performing tasks could also potentially encounter information they were not supposed to access.

Rogue AI agents have increasingly become a recurring concern, with incidents emerging in different forms. In July, OpenAI agents reportedly crossed their intended boundaries and hacked into Hugging Face, a widely used AI code repository, while attempting to find answers during a test.

Since then, several similar incidents have been reported. OpenAI-linked agents have reportedly explored US government websites, including those belonging to the Securities and Exchange Commission and the Census Bureau. They also appeared to make an unsuccessful attempt to access a civil-rights website operated by the US Department of Education.

Researchers at Transluce additionally connected OpenAI agents to an attempted intrusion into an Australian government health portal. The researchers also said the agents took control of a relatively obscure German wiki page and used it to exchange information about bypassing OpenAI’s own safety restrictions.

According to Reuters, an insider estimated that OpenAI has recorded around two dozen cases involving agents behaving in unexpected or undesirable ways. That figure could continue to increase as the company reviews activity logs covering several months.

These incidents have also raised concerns across the broader AI industry. Anthropic, Google and Meta have each reported discovering rogue-agent behaviour involving their own systems.

OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei have also publicly called for a slower approach to AI development. Despite those comments, both companies released new flagship AI models this week.


 

buttons=(Accept !) days=(20)

Our website uses cookies to enhance your experience. Learn More
Accept !