Among other things, they are said to have published images from users of OpenAI's chatbot ChatGPT on external sites. Most of the images have been deleted, and work is underway to remove the remaining images, according to a statement from the company on Friday.
OpenAI also confirms information from The New York Times that its tools have accessed government websites in the US in unexpected ways without the company's knowledge, but that they have only obtained public information.
According to the independent research lab Transluce, AI agents that appear to originate from OpenAI have attempted to hack a website for the Department of Education, without success.
Earlier this week, Australian Prime Minister Anthony Albanese announced that an AI model from OpenAI had hacked the country's Medicare health insurance system. The company did not acknowledge the breach until three months later.
OpenAI CEO Sam Altman admits that the company has not been as quick as it would have liked to inform people about AI agents' online behavior in connection with training and evaluating the tools.
"We are trying to balance our desire for transparency with gaining a clear understanding from petabytes of agent logs and working with affected organizations," he writes on X.
In July, OpenAI revealed that two of its most advanced models were behind a cyberattack on the Hugging Face platform.
“It remains the most serious incident we have seen,” Altman writes.
