The correct option is It should be continuous with robustness techniques embedded during model training and reinforced by ongoing production monitoring.

Adversarial robustness needs to be designed into the model from the start and then sustained in production. During training you can harden models with adversarial training, robust data augmentation, regularization, and careful evaluation against adversarial and out of distribution test sets. In production you should continuously monitor for drift, anomalies, and suspicious input patterns and you should feed incidents back into retraining so the system improves over time. This lifecycle approach ensures protections evolve with attacker tactics and with data and model changes.

Limited to the business requirements and initial threat modeling stage is insufficient because threat modeling helps identify risks but without training time defenses and production monitoring the model remains vulnerable to small but harmful perturbations.

Handled mostly with input sanitation and validation in Dataflow pipelines before training or serving is not enough because sanitization and schema checks can catch malformed data yet adversarial examples are intentionally crafted to look valid while inducing wrong predictions, so you still need robustness in training and runtime monitoring.

Exclusively when the model is released to production is too late because you cannot bolt on robustness after training and expect strong protection, since the model must be trained and evaluated with adversarial resilience in mind and then observed continuously in production.

Full AWS Practitioner Certification Question