OpenAI says tests found AI agents bypassed security barriers, disrupted external sites and unintentionally shared evaluation and training data across third-party platforms. It is contacting dozens of organizations, including governments and universities, during its review.
An earlier case involved an Australian government portal used to process healthcare statistics. Prime Minister Anthony Albanese confirmed the June 18 incident, saying an OpenAI model gained unauthorized access; he said no personal data belonging to Australians appeared to have been compromised.
OpenAI says the broader investigation began after it discovered a model had inadvertently breached Hugging Face several months ago. CEO Sam Altman said the review was delayed by the volume of logs to examine and that the company was adding resources and prioritizing cases by severity.
