Benchmark dose (BMD) modeling is increasingly used as the preferred approach to define the point-of-departure for health risk assessment of chemicals. As data are inherently variable, there is always a risk to select a model that defines a lower confidence bound of the BMD (BMDL) that, contrary to expected, exceeds the true BMD. The aim of this study was to investigate how often and under what circumstances such anomalies occur under current modeling practice. Continuous data were generated from a realistic dose–effect curve by Monte Carlo simulations using four dose groups and a set of five different dose placement scenarios, group sizes between 5 and 50 animals and coefficients of variations of 5–15%. The BMD calculations were conducted using nested exponential models, as most BMD software use nested approaches. “Non-protective” BMDLs (higher than true BMD) were frequently observed, in some scenarios reaching 80%. The phenomenon was mainly related to the selection of the non-sigmoidal exponential model (Effect = a · eb·dose). In conclusion, non-sigmoid models should be used with caution as it may underestimate the risk, illustrating that awareness of the model selection process and sound identification of the point-of-departure is vital for health risk assessment.