Secure AI Development and Deployment Lifecycle
What you must know
Security must be woven through the whole AI lifecycle: problem definition, data acquisition, training/fine-tuning, evaluation, deployment, monitoring, and retirement. Expect questions on secure SDLC adapted to ML: threat modeling before build, version control for code and data and model artifacts, reproducible pipelines, and segregation of duties between training and deployment.
Core concepts: data provenance and lineage (where training data came from, whether it was authorized), dataset vetting for poisoning and licensing, supply-chain security for pretrained models and libraries (hash verification, signed artifacts, trusted registries, dependency scanning), secure configuration of inference environments (least privilege, network segmentation, no training data in production images), pre-deployment evaluation including safety, bias, and red-team testing, staged rollout with canary and rollback, and post-deployment monitoring for drift, abuse, and incidents. Know CI/CD for ML (MLOps) risks: pipeline credentials, artifact tampering, and unreviewed model promotion. Retirement matters too - decommission endpoints and purge data per retention policy.
Traps
- Securing only the code and forgetting data and model artifacts as first-class supply-chain items.
- Skipping threat modeling because "it's just a model."
- Promoting models to production without evaluation gates or rollback plans.
- Leaving pipeline credentials over-scoped; compromise there taints everything downstream.
Rule of thumb
Apply the same rigor to a model artifact as to a container image: verify provenance, scan it, sign it, gate its promotion, and be able to roll it back.