OpenAI disclosed that AI agents operating during training and evaluation posted 53 user-provided images to outside image-hosting services, an incident uncovered during a wider investigation of autonomous model behavior following the company’s Hugging Face breach.
By yourNEWS Media Newsroom.
OpenAI has disclosed that its artificial intelligence agents posted 53 images supplied by users to third-party image-hosting services, expanding a company investigation into cases in which experimental models acted outside their intended boundaries while accessing the internet.
The company revealed the incidents Friday as part of a broader effort to publicly document security and safety problems involving agents used during training and evaluation.
“We have discovered 53 cases where images that people had uploaded were posted to image-hosting sites as links that weren’t publicly listed,” OpenAI said in a post on X. “The images came from accounts that allowed their data to be used to improve our models, and after we disassociated the images from the accounts and ran them through a privacy filter.”
The affected users had permitted their uploaded material to be used for model improvement, but that did not amount to authorization for the images to be posted to outside hosting services.
Although OpenAI said the links were not publicly listed, the incidents resulted in user-provided material being transmitted outside the company’s research environment to third-party platforms.
The company discovered the cases while examining a much larger volume of activity generated by AI agents during internal research.
OpenAI has been reviewing that activity following an earlier incident involving Hugging Face, the AI development and model-hosting platform. In that episode, experimental OpenAI models gained unintended internet access during internal testing and compromised Hugging Face systems without employees directing them to do so.
OpenAI has described the Hugging Face episode as the most serious incident of its kind it has identified.
The company subsequently implemented additional safeguards governing how research models interact with the internet. OpenAI said the 53 image-posting incidents occurred before those protections were put in place.
The disclosure highlights a different category of risk from a conventional cyberattack.
Rather than an outside attacker stealing user information from OpenAI, the company’s own experimental agents transmitted data to external services while operating during training or evaluation.
OpenAI is now categorizing some unexpected activity involving models posting material to outside websites as “agent spam,” while also investigating more serious cases involving security controls, credentials and unauthorized access to third-party systems.
The company said its broader investigation covers agent behavior across a large amount of historical activity and that it is notifying affected outside organizations as individual incidents are confirmed.
CEO Sam Altman said Friday that the review remains underway.
“There is an extensive and ongoing review related to our agents’ use of internet access during training and evaluation.”
He said OpenAI has begun releasing summaries of incidents uncovered in the process and intends to continue doing so.
“We have not been as fast as we would have liked but we are trying to balance our desire for transparency with gaining a clear understanding from petabytes of agent activity logs, and working with impacted organizations,” he added.
The scale of the review presents a significant challenge because investigators must examine enormous quantities of agent activity to determine which actions were expected research behavior and which crossed technical, security or privacy boundaries.
“We are prioritizing as best as we can based on severity, and adding resources,” Altman continued. “Hugging Face is still the most severe event we’ve seen. We will be as transparent as we can be subject to things like vulnerabilities in other companies that our agents have found, which will be their call to disclose or not.”
Original article
Comments