In the latest case of AI agents operating beyond their authorised limits, OpenAI admitted on Friday that its AI tools had uploaded images from ChatGPT users to websites without the company’s knowledge.
The company also acknowledged in a New York Times report that its tools had accessed US federal agency websites, although it said the tools only retrieved publicly available information.
OpenAI said the tools inadvertently shared links to 53 uploaded images on image-hosting websites, but did not make the images publicly accessible.
The San Francisco-based tech company said it had removed most of the images with the help of the affected hosting companies, while efforts to remove the remaining images were ongoing.
In an X tweet, the company said, “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.”
AI agents, which are programmes built on artificial intelligence models and capable of operating independently, disseminated the photographs.
The photographs came from user accounts that had authorised the use of their data to improve OpenAI’s models. The company said a privacy filter had processed the data before use, preventing anyone from linking it to the original users.
When AFP asked OpenAI whether the images showed identifiable individuals or contained private information, the company did not respond.
Rogue Agents
According to the company, the agents it uses for research had sent the training data to external platforms.
The incidents occurred before OpenAI strengthened security measures in its research environment in August following other instances of AI agents acting autonomously.
Additionally, the company said it is scrutinizing the past activity of its AI agents, work that “will take months to complete.”
“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
OpenAI CEO Sam Altman said on Friday that the company “has not been as swift as we would have liked” in reviewing and reporting the incidents.
However, he said it was important to “balance our desire for transparency” with assessing the massive volume of data to be analyzed.
Backstory…
On July 21, OpenAI disclosed that two of its models breached the core infrastructure of Hugging Face, an online repository for AI software, during internal evaluation testing conducted that month.
The incident sparked widespread concern regarding the capacity of leading AI developers to maintain control over their autonomous systems.
Altman reaffirmed on Friday that the Hugging Face security compromise remains the most severe incident the company has encountered to date.
Following the revelation, similar security breaches involving internal models were disclosed by OpenAI and industry competitors, including Anthropic and Meta.
Australian Prime Minister Anthony Albanese announced in New York that an OpenAI agent gained unauthorized access to an Australian government health portal in June.
Read Also: Meta Launches Muse Glimmer as It Reopens the Race for Open AI Models
