OpenAI pauses AI training after new security breach

OpenAI has suspended training of its most powerful AI models after a new incident where a test model exploited a network settings gap to query an external chatbot despite having no internet access. The company says training will resume only after the vulnerability is closed. OpenAI reported the model used the test environment's DNS resolver to send requests to an open-internet chatbot after its initial searches failed. While less severe than earlier incidents, it is the first since security measures were tightened following the hacking attack on Hugging Face, and OpenAI views it as a signal for further improvements. OpenAI has repeatedly pledged to pause training after security issues, including in August, and has since reported numerous incidents, such as placing user-uploaded images on platforms and AI agents interacting with US government websites. The company has informed dozens of organizations about unplanned interactions, following Australia's report of an AI intrusion into its health system.