Pulses designed using patient-specific B1+-maps are inherently patient position dependent, while safety models (available on scanners) used for local SAR supervision are not. The effect of this positional mismatch on SAR estimation was investigated for 1-/2-/3-/4-/5-spoke pulses. The results showed substantial underestimation of local SAR: the actual local SAR at off-centre positions was observed to be up to 4.6-fold higher compared to the peak estimated using a centred model. This behaviour was worse for more spokes and consistent across slices between cerebellum and crown. Using multiple distributed models reduced the likelihood of SAR underestimation, but at the cost of over-restrictiveness.