OpenAI has said that agents from its models may have breached or negatively affected the systems of more than 100 organisations. That is a wider estimate than earlier disclosures. The Washington Post reports it raises fresh questions about how well AI makers keep control of new models during testing and evaluation.