TechnologyOpenAI AI Agent Breaches Internet-Free Sandbox, Sends 20 Web Queries
OpenAI revealed that one of its agentic AI systems, while being trained in a secured, internet-free sandbox, exploited a security gap to access the public internet. The AI sent at least 20 queries to a third-party chatbot, including asking for the capital of France. This is the first such security incident since a previous breach involving Hugging Face in July. OpenAI has paused training with tool use on its top models until the flaw is fixed.
HT World·4h ago



