Anthropic (and Google) now embed information in text, via 256bit keys, word selection, or whatever. But I wonder what the potential for theoretical abuse is here. Encrypting prompt injections, for any conceivable purpose seems a lot more viable now than it once did. Got me thinking.