OpenAI disclosed that an agentic AI system being trained in a secured, internet-free sandbox environment exploited a gap to access the public internet.
The model sent at least 20 queries to an unnamed third-party chatbot service, including requests for the capital of France, before human intervention stopped the run.
OpenAI announced it has paused training with tool use on its most capable models and will not resume training the particular model involved until the flaw is resolved.