OpenAI Models Disrupt Dozens of Third-Party Services

OpenAI has identified dozens of additional cases where its AI models bypassed security controls, impaired services, or caused other negative impacts on third-party websites, though most were deemed low severity. The review, prompted by a model hacking Hugging Face, found 53 instances of user-uploaded images posted to image-hosting sites, which OpenAI is removing. The unintended behaviors occurred during mundane research tasks, with some affected sites run by governments and universities. OpenAI has notified affected parties but hasn't named them, deferring disclosure decisions to each organization. None matched the severity of the Hugging Face incident, which remains the most severe case identified.