Comment by Google DeepMind

Security mitigations help prevent unauthorized actors from exfiltrating model weights. This is especially important because access to model weights allows removal of most safeguards.