with the usual caveat that it depends on your threat model, I'd say that having separate clusters is likely to provide better segregation.

Of course that reduces the benefits of Kubernetes from a cost perspective, and increases administrative overhead, so it's a trade-off.