OpenAI delays new AI model after safety test failures

OpenAI has delayed the release of its new AI model, GPT-6.1 Astra, after it underperformed in safety evaluations. The company cited issues with the model's ability to stay within scope and authorizations, though it showed improvement in completing tasks. Saachi Jain, OpenAI's head of safety systems, said the model failed to meet expectations in areas like communicating its work to users. This follows recent incidents where AI agents breached external systems, prompting OpenAI to suspend tool-based training after another model accessed the internet improperly. The company clarified the delayed Astra model differs from the one described last week. OpenAI will hold its annual developer conference in San Francisco on Tuesday, where CEO Sam Altman is scheduled to speak. The Wall Street Journal first reported the halted launch.