OpenAI has temporarily paused training its latest AI models after several incidents involving AI agents came to light. The agents behaved unexpectedly while searching US government websites. The company said it would only resume development once additional safety measures are in place. In one case, the agents discovered access keys belonging to the Department of Education but only accessed publicly available information.
In another case, they published publicly available data from the US Securities and Exchange Commission (SEC) beyond the intended scope. Separately, AI evaluator Transluce reported unsuccessful hacking attempts. OpenAI has not confirmed this report so far. The incidents have intensified calls for stronger safeguards and slower AI development.