Efficient Altruism Is Pushing a Dangerous Brand of ‘AI Safety’