Deep learning applied to bioinformatics pipelines has allowed predictions that were thought out of reach; it has also introduced another source of cyber-biosecurity threats to deal with. The present work offers a new framework that can examine and improve the adversarial robustness of AI-based splice site detection models. The starting point is a binary-encoded genomic data: from this an exon–intron (EI), intron–exon (IE), and non-boundary sequences are re-constructed and is used to train a long short-term memory (LSTM) model to perform the above classification process. Simulation of adversarial settings is carried out by performing fine grained perturbations in the nucleotide level that interferes with the predictions in a biologically plausible manner to reduce sensitivity against critical boundary sites of the model. To counteract this weakness, an adversarial training approach is taken whereby perturbed series are progressively injected into the learning process in which the model becomes capable of generalizing on both unmodified and corrupted data. Theoretical findings found that splice site prediction models are naturally delicate to even the slightest manipulation of its inputs; yet, with controlled immersion of adversarial examples, sufficient robustness may be proficiently retrieved. The research provides a new interpretable and reproducible style of adversarially resilient learning to the sequence-based bioinformatics domain, thus promoting AI safety and cyber-biosecurity.