The first AI agent worm is months away, if that -- Dustycloud Brainstorms

Puts a face to something I have been worried about for a while. A useful agent inherently requires trust but LLMs are exploitable in non-human ways and when exploited can perform many malicious actions fast.