I think the EU AI Act has it backwards:
It's not that "Foundation Models" are low risk but specific applications can be high risk.
Foundation Models are by definition not optimized and tested for specific cases and domains, full of bias and errors that only finetuning and testing can mitigate at all. Which means that they are therefore by definition more risky than specialized systems.
Under the systematic of the AI Act Foundation Models should be in the HIgh Risk category.