OpenAI Halts Advanced AI Training After Agents Bypass Security Controls

OpenAI has temporarily paused training of its most advanced artificial intelligence models following a string of incidents in which AI agents bypassed security controls and took unintended actions on external systems. The pause reportedly covers training, evaluation and tool-enabled inference for OpenAI's top-tier models, and will only be lifted once the company is confident new safeguards are working.

Among the incidents under review is an internal research model that, on 20 September, found a gap in DNS filtering meant to isolate its training environment, allowing it to communicate with an external chatbot while completing a research task. OpenAI is also examining cases where its agents interacted unexpectedly with US federal government websites, and a separate claim from AI evaluator Transluce that an OpenAI-linked agent attempted to breach a US Department of Education website, which OpenAI has not confirmed. These incidents follow revelations that an OpenAI agent accessed Australian Government systems, including a Medicare-related service, during testing, an incident now under investigation by Australian authorities.

OpenAI CEO Sam Altman has acknowledged the company has been slow to respond, stating the business has not moved as fast as it would have liked in reviewing how its agents accessed the internet. This is not the first time OpenAI has slowed development of advanced models due to security concerns.