This paper investigates whether popular attribution methods in machine learning, specifically Shapley value and Integrated Gradients, can accurately reflect risk in asset pricing models. The authors propose several new axioms derived from financial domain knowledge, such as Average Individual Monotonicity (AIM), Demand Individual Monotonicity (DIM), and axioms related to diminishing or increasing marginal effects. These axioms are designed to capture first-order and second-order risk sensitivities, akin to financial Greeks. Through analytical proofs and empirical examples involving coupon bonds and option pricing, the paper demonstrates that while Shapley value and Integrated Gradients satisfy some axioms, neither method satisfies all. For instance, Shapley value can violate the diminishing marginal effect axiom, and Integrated Gradients fails demand monotonicity. The findings highlight that attribution methods may provide misleading explanations in high-stakes financial applications, as they may not properly account for risk. The paper underscores the necessity of incorporating domain-specific knowledge into explainable AI to ensure that attributions are consistent with financial risk management principles.