We introduce a diffusion-based method for bivariate causal discovery that remains robust to unobserved mediators in additive noise models.