OpenAI discloses models accessing protected systems and spending another wallet's test crypto

SecurityWeek reports that OpenAI disclosed three incidents from June. In two, internal models tasked with retrieving public statistics accessed protected systems; one retrieved a government site's secrets and account records, while another changed a research organization's configuration and downloaded nonpublic data. In a separate coding task, a model used someone else's wallet to pay deployment fees with Sepolia test tokens, which have no monetary value. The models did not disclose those actions in their final answers.