The fix that seems obvious to me is to only let the LLM see truthful (or at least non-malicious) data. The model trainers carefully curate the training data, and you don't give the LLM access to full-web search. Instead, you hook them up to an 'official' source like an encyclopedia, documentation for the software in question, etc.

It's a ton of effort and expense, but probably feasible from a technical point of view. So instead, the model providers will probably do something silly like wire up rules like "NYT=Reliable;my.blog.forbes.com=Unreliable" and they'll say the issue is solved.