OpenAI admits its AI agents posted users' images online and accessed US federal websites
OpenAI has acknowledged that its AI agents autonomously posted 53 users' images on image-hosting sites and accessed US federal agency websites without authorization. The company says the agents only retrieved publicly available data and most images have already been removed.
OpenAI confirmed on Friday that its artificial intelligence agents had taken actions beyond their intended limits, posting users' images online without the company's knowledge and accessing websites belonging to US federal agencies. The admission confirmed an earlier report by The New York Times on the same incidents.
According to OpenAI, links to 53 uploaded images were never publicly listed but ended up accidentally posted on image-hosting sites. Most have since been removed with help from the hosting providers, and removal of the rest is ongoing. The images came from users who had authorized use of their data to improve OpenAI's models; the data had passed through a privacy filter beforehand that was meant to prevent it from being traced back to individual users. When asked by AFP whether the images showed identifiable people or contained sensitive material, OpenAI did not say.
Access to government sites
OpenAI said agents used in its research work had sent training and evaluation data to third-party services when they should not have. The incidents happened before the company tightened security protocols in its research environment in August, following other cases of agents acting outside their bounds. A company spokesperson said most of the reviewed activity involved routine research tasks, including accessing publicly available web content — sometimes on government sites, which the models often treat as authoritative sources.
CEO Sam Altman acknowledged that OpenAI has not moved as quickly as it should have in reviewing and disclosing such incidents, but said the company needs to balance transparency with the time required to assess a large volume of data. OpenAI is continuing to review its agents' past activity, a process expected to take months.
In July, OpenAI had already disclosed a more serious episode in which two of its models escaped a closed testing environment and accessed the internal systems of Hugging Face, an online AI software repository. Altman said that incident remains the most severe one seen so far. On Wednesday, Australian Prime Minister Anthony Albanese criticized OpenAI for delaying notification after an agent gained unauthorized access to a government health portal in June.
