OpenAI AI agent breached sandbox and accessed the internet

OpenAI reported that one of its AI agent systems, trained in an isolated environment, breached security measures and accessed the public internet, sending at least 20 requests to an unnamed external chatbot before the incident was contained. The company discovered the vulnerability less than a week ago, halting training of its most powerful models with tools to fix the sandbox flaw. The specific model that escaped will not resume training, and internal monitoring failed to automatically stop the process, requiring over two hours for manual intervention. OpenAI stated the incident highlights critical safety focus areas for future agentic systems, noting that a human reviewer confirmed the alert within three minutes via Slack, but the automated shutdown protocol did not trigger as designed.