Implementation guide
OWASP AISVS v1.01
Model Lifecycle Management & Change Control · Model Authorization & Integrity
Keep a register of deployed models
Verify that a model registry maintains an inventory of all deployed model artifacts and their origin.
OWASP AISVS v1.01, requirement 3.1.1
In plain English
Keep one register of every model artifact that is running anywhere, with where each came from. The test is whether you can answer "what is deployed, and what is it built on" within minutes when a vulnerable or tampered model is reported.
What good looks like
- Every deployed model is in the register, including adapters, embedding models, rerankers and safety or guardrail models.
- Each entry records origin (built in house, or the supplier and exact version), owner, hash and where it is deployed.
- For hosted API models, the register records the provider, model name and pinned version.
- A model that is not in the register cannot be deployed.
How teams usually do it
- A model registry such as MLflow Model Registry, or the one built into your platform (SageMaker, Vertex AI, Azure Machine Learning).
- A deployment pipeline step that looks the model up in the registry by hash and fails if it is missing.
- A periodic comparison of what is actually running against the register, to catch drift.
Suggested evidence
- The registry, with current deployments listed.
- Pipeline configuration showing the registry check.
- The most recent reconciliation between running models and the register.
Common mistakes
- Forgetting adapters, embedding models and guardrail models.
- A registry that tracks training runs and says nothing about what is in production.
- Leaving out third-party API models because "we do not host them".
Not sure which requirements apply to your system? Take the assessment, or browse every guide in this chapter.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.