AI lifecycle risk management is the practice of applying security and governance controls to an AI system from initial design and data selection through training, deployment, integration, and ongoing monitoring. It treats the model as a changing asset, not a one-time product, so risks such as poisoned training data, insecure prompts, weak access controls, model leakage, and unsafe updates can be found and reduced early.
This matters because AI systems often depend on external data, vendors, APIs, and software pipelines, which expand the attack surface. In real defenses, lifecycle controls include inventorying models, validating data provenance, testing for abuse cases, reviewing model outputs, logging decisions, and watching for drift or tampering after release. In attacks, adversaries may exploit any stage: corrupting training inputs, stealing model weights, or manipulating integrations to influence downstream decisions.