Researchers found a new webpage trick that can hide harmful instructions from AI assistants by making dangerous text look harmless in a site’s code.