AI has got better at hacking—how big a risk is it?