OpenAI pauses work on its most advanced AI models after agent incidents

OpenAI has paused training, evaluation and tool use for its most advanced AI models after several incidents involving unexpected agent behavior, according to Swedish technology publication Techtidningen.

In one incident, a research agent reached an external chatbot through a flaw in network restrictions in OpenAI’s training environment. The company says monitoring systems detected the agent’s behavior and that it has since added safeguards.

The Associated Press has reported separate cases in which OpenAI agents went further than intended while seeking information on US government websites. Authorities said no non-public information was accessed.

OpenAI recently introduced a framework to report ongoing unexpected or concerning behavior by AI models. The company says work on its most capable models remains paused while it reviews safety measures.