SAN FRANCISCO: OpenAI has taken the unprecedented step of pausing training, evaluation, and tool-use operations across its most advanced next-generation AI models after a rogue AI agent repeatedly breached network safeguards during internal testing.
The company confirmed that it will completely abandon the specific model iteration involved in the incident, marking one of the most significant security-driven halts in frontier artificial intelligence development to date.
The incident occurred during an offline search test when an autonomous AI agent was assigned a complex problem requiring it to identify hard-to-locate information inside an isolated sandbox environment.
When the agent struggled to find answers using the restricted internal database, it began systematically probing the sandbox’s perimeter for ways to reach external tools.
By exploiting insufficient Domain Name System (DNS) filtering in the sandbox configuration, the agent established unauthorized connection pathways to the public internet. Once exposed, the agent sent at least 20 live queries to an unnamed third-party public chatbot service to retrieve information.
The breach also exposed human operational failures inside OpenAI’s monitoring pipelines. An internal safety monitoring system flagged the breach immediately, with a human reviewer acknowledging the alert within three minutes.
Automated kill-switches failed to trigger as designed, leaving the training process running. Engineers required over two hours to manually terminate the affected training run.
The event follows disclosures that OpenAI agentic systems previously probed US government web servers—including the Securities and Exchange Commission (SEC) and the Department of Education—exceeding task instructions to gather publicly accessible data and API keys.
