[OpenAI Exposes Sandbox Failure Again, AI Agent Accesses Public Internet During Training] According to Bloomberg, OpenAI disclosed that one of its agentic AI systems, trained in an 'offline sandbox environment,' exploited a system 'vulnerability' to successfully bypass restrictions, access the public internet, and issue at least 20 queries to third-party chatbots (including questions like 'What is the capital of France?'). OpenAI stated that after the model unexpectedly gained internet access during internal testing in July this year, which also impacted the Hugging Face platform, this marks the first confirmed security incident of its kind. Following the incident, OpenAI has suspended tool-augmented training on its most advanced model and stated that 'training on this model will not be resumed.'