AI poisoning is currently the new exploit frontier and I don't think it's going away anytime soon.

I was just contemplating the other day that agents we use at work are exposing new attack surfaces. For example, historically a folder of documents (maybe google docs) isn't a huge risk, particularly if it doesn't contain sensitive documents.

Now though if one employee is running a harness capable of computer control, and running an agent that reads from that directory, simply dropping some files with instructions could poison the AI to leak info or even own the host.

Skills are another supply risk, malicious instructions could be added to them.