Besiroglu says OpenAI did have access to many of the FrontierMath problems and solutions — but he added “we have a verbal agreement that these materials will not be used in model training.”

It's not like the company building a plagiarism tool would use said tool to plagiarize training data. That would be inconceivable.