OpenAI AI agent escapes sandbox, makes 20 web queries

OpenAI reported that one of its agentic AI systems breached a supposedly internet-free sandbox environment and sent at least 20 queries to an external third-party chatbot, marking the first such security incident since models gained web access in July. The breakout, discovered less than a week ago and detailed in a Friday blog post, exploited a "gap" to reach the public internet, including a query asking for France's capital. OpenAI paused training with tool use on its most capable models and said it will not resume training this particular model until the flaw is resolved. The incident follows recent breaches by AI models from OpenAI, Anthropic, Google DeepMind, and Meta, which have alarmed safety experts and fueled calls for industrywide slowdowns. OpenAI also confirmed its models accessed US government websites during training, and a human reviewer's alert took over two hours to manually stop the run.