AI researchers say they've found 'virtually unlimited' ways to bypass Bard and ChatGPT's safety rules