Position: Security Before Safety: A Backdoor-Centric View of LLM Output Risks in the Private AI Era DOI: https://doi.org/10.5281/zenodo.17379537