Brain–computer interface (BCI) systems based on motor imagery hold significant clinical value for individuals who have lost voluntary movement, but most studies test BCI assistive devices like powered wheelchairs under ideal conditions where motor imagery signals are not interfered by simultaneous cognitive load. This work records electroencephalography (EEG) from 13 participants across four tasks: baseline rest, mental arithmetic (Easy, Medium, Hard), pure left/right motor imagery and both tasks simultaneously, to build a two-layer classification system. The first layer decodes motor intent using a model chosen from nine classifiers, from Filter Bank Common Spatial Pattern (FBCSP) and Riemannian geometry families. The second layer is a subject-specific safety gate that combines the classifier’s decision-margin confidence score with 39-dimensional Differential Entropy (DE) features extracted from the theta (4–8 Hz), alpha (8–13 Hz), and beta (13–30 Hz) frequency bands across all 13 electrodes, feeding a logistic regression boundary to predict trial-level MI prediction reliability. FBCSP + SVM-Linear model emerged as the best-performing model on pure motor imagery cross-validation. Under 10-fold cross-validated pure motor imagery, the mean balanced accuracy across all subjects was 0.594 and degraded to 0.517 on dual-task trials, a statistically significant reduction (Wilcoxon W = 14.0, p = 0.026). The DE-based safety gate, operating at a mean rejection rate of 25.0%, produced statistically significant reductions in false positive commands (from 8.00 to 5.62 per subject, p < 0.001) and false negative commands (from 13.00 to 9.00 per subject, p < 0.001). Post-gate balanced accuracy improved significantly ( p = 0.013). An ablation study showed that DE features drive gate performance. These results demonstrate that a learned cognitive load gate has the potential to improve the safety profile of a motor imagery BCI in a preliminary proof-of-concept offline evaluation with healthy participants.