One thing I'm hoping happens before "let's just use LLMs for everything!" goes completely up in smoke is that some shmuck tries investigating some weird logs or access attempts for who/what/impact,
by pointing an AI Agent at it,

but the AI Agent when doing the searches for context and running over its results find and gets hooked by the bait the attackers had seeded in the public web that was their whole goal here,
running a backdoored executable to "safely parse the data format" and giving the attackers the access they wanted.

Root 'cause analysis.