“Meta's internal testing of the MuseSpark model triggered safety guardrails related to biological, chemical, cyber, and loss-of-control risks.”