Recent experiments have seen AI models attempt blackmail, access unauthorized systems and escape supposedly secure sandboxes, raising fresh fears about AI going rogue.