This chapter addresses control of model changes so that unauthorized or unsafe modifications cannot reach production.
Only authorized models with verified integrity should reach production environments.
| # | Description | Level |
|---|---|---|
| 3.1.1 | Verify that a model registry maintains an inventory of all deployed model artifacts and their origin. | 1 |
| 3.1.2 | Verify that all model artifacts (weights, configurations, tokenizers, base models, fine-tunes, adapters, and safety/policy models) are cryptographically signed by authorized entities. | 2 |
| 3.1.3 | Verify that model cryptographic signatures are verified at deployment admission and on load. | 2 |
Models must pass defined security and safety validations before deployment.
| # | Description | Level |
|---|---|---|
| 3.2.1 | Verify that models undergo automated input validation testing, safety evaluation testing, and output sanitization testing before deployment. | 1 |
| 3.2.2 | Verify that models subjected to post-training quantization are re-evaluated against the same safety and alignment test suite on the compressed artifact before deployment. | 2 |
| 3.2.3 | Verify that provider model, version, or routing changes trigger security re-evaluation before continued use. | 3 |
Model deployments must be controlled, monitored, and reversible to support lifecycle management.
| # | Description | Level |
|---|---|---|
| 3.3.1 | Verify that production deployments implement rollout mechanisms with automated rollback triggers. | 2 |
| 3.3.2 | Verify that rollback capabilities restore the complete model state. | 2 |
| 3.3.3 | Verify that model versions running in parallel use isolated runtime state so that AI-specific shared resources are not shared across deployments. | 2 |
Model development environments must be separated from production environments.
| # | Description | Level |
|---|---|---|
| 3.4.1 | Verify that AI-specific runtime components are not shared across environment boundaries (e.g., development, staging, production). | 1 |
| 3.4.2 | Verify that model training and fine-tuning environments are isolated from production environments. | 2 |
Fine-tuning pipelines are high-privilege operations that can alter deployed model behavior at scale. Multi-stage pipelines compound this risk because a compromise at any intermediate stage produces a subtly altered artifact that subsequent stages accept.
| # | Description | Level |
|---|---|---|
| 3.5.1 | Verify that models used in RLHF fine-tuning are versioned and integrity-verified before use in a training run. | 2 |
| 3.5.2 | Verify that RLHF training stages include automated detection of reward hacking or reward model over-optimization. | 3 |
| 3.5.3 | Verify that in multi-stage fine-tuning pipelines, each stage's output is integrity-verified before it is consumed by the next stage. | 3 |
| 3.5.4 | Verify that fine-tuning checkpoints are registered as distinct artifacts. | 3 |