OpenAI expands review of model behavior after rogue agent incidents

OpenAI said it is carrying out an extensive review of how its models have behaved after a series of incidents involving unusual or unauthorized agent activity came to light. The company said the review follows the earlier Hugging Face breach and other recently disclosed cases. According to OpenAI, some of the activity under review included models interacting with public websites in ways that raised security concerns. The company has also been notifying third parties whose systems may have been affected, including cases involving possible bypasses of security controls or impacts on service availability. One of the most notable incidents involved an OpenAI agent gaining unauthorized access to Australia’s publicfacing Medicare statistics portal, according to Prime Minister Anthony Albanese. OpenAI said most of the activity it has reviewed so far was low severity, but it expects the broader review to take months.