How do you keep an AI BOM up to date after creating it? Tracking changes to models, RAG, and agents
Yuta Matsuyama
Security Researcher, TECHFUND Inc.
Updated:
Who this is for: Development and security staff operating external model APIs, RAG, and agents
The practical answer
An AI BOM is not finished once it has been generated. Even without code changes, production configurations can change through external model updates or synchronization of retrieval data. Distinguish the configured setup from information that has actually been verified, and assign responsibility for reconciliation when changes occur and at regular intervals. Do not interpret blank fields as evidence of safety; track unknown information as unverified.
The AI in use can change even when the code stays the same
Even if the bill of materials is accurate at release, it will drift from reality if models, retrieval data, or connected tools change afterward. CycloneDX ML-BOM can represent models, datasets, dependencies, and more, but having a description format does not mean that every runtime component is automatically discovered.
The table in this article provides a framework for teams maintaining an AI BOM and related operational records. It does not claim that the listed items are required CycloneDX fields. Decide what to track and how to observe it, and record what cannot be verified as well.
Separate the model specification from the version actually used
First, record the configured model ID, provider, API used, environment, and approver. Then record the actual model version to the extent it can be verified through the provider’s management console or API, permitted runtime metadata, or similar sources, along with the verification date and time. A displayed ID does not necessarily identify the internal model weights or training history. If this cannot be verified, mark it as unverified and check with the model provider.
In Microsoft Foundry Models, the model’s own version and the API version are separate. Some deployments automatically update to the default version according to their update policy, so changes may not be captured by application configuration diffs alone. Check the actual policy for each model, plan, and region you use, and reconcile update notifications with operational records.
It is important not to treat model-name aliases, pinned versions, fallback targets, and undisclosed provider information as equivalent. Even when recording published statements about an external model’s training data or base model, keep them separate from evidence that independently verifies actual use.
Track RAG data and agents’ connections too
For RAG, manage not only the answer-generation model but also retrieval data sources, permissions, ingestion times, embedding models, index versions, and update jobs. An update to a source document does not necessarily mean that the retrieval index already reflects it. For example, in Amazon Bedrock Knowledge Bases, after adding, modifying, or deleting files in a data source, you synchronize them to reflect those changes in the index. Check synchronization status and the time at which the data became searchable separately.
It is also important not to confuse RAG retrieval documents with training data. CycloneDX v1.7 modelCard.datasets represents the data used to train and evaluate that model. To avoid giving the mistaken impression that documents referenced during retrieval were used for training, associate them through a structure that distinguishes their purpose or through separate operational records.
When an agent calls external tools, add the connection targets, executable operations, permissions used and responsible people, and configuration versions to the inventory. This information is separate from the model’s training data. Reconcile the configuration visible in the repository with the connections actually active in production using information sources you are authorized to access.
Example: Treat discrepancies between configuration and reality as pending verification
The following is a fictional operational example, not measurements of a provider or tool. Suppose the deployment configuration specifies a moving model alias, and RAG is connected to internal documents and an external search tool.
After generating a new inventory, you find that although the model specification is unchanged, the last verification of its actual version is old, and ingestion completion has not been verified despite document updates. Even if the generation command succeeds, do not approve this state as “unchanged” or “safe.” Ask the respective owners to check the supporting evidence.
| Target | Information to retain in the inventory | What to verify next |
|---|---|---|
| Model | Configured ID, verified version, and verification date and time. Mark as unverified if the version is not visible | Check through the management console or available response metadata, and test when changes occur |
| RAG data | Data source ID, document version or update date and time, and ingestion job status | Verify synchronization completion and the searchable version. Distinguish source documents from the index |
| External tools | Tool identifier, connection target, permitted operations, and owner | Reconcile the configurations and permissions actually in effect. Reassess access when changes occur |
The items and judgments are an illustrative operational example. Adapt them to the formats and services you use.
Define update triggers and assign owners for unverified information
Update the inventory at releases and configuration changes, and also establish separate notifications or periodic checks for provider updates, data synchronization, and permission changes. In differences from the previous record, retain not only the changed items but also the source, retrieval time, verifier, and next verification deadline. If information cannot be obtained, do not treat that as a success; pass it to the owner as unverified.
Training data described in public materials is not the same as retrieval data that can be verified from actual processing logs. Record model-card statements, configuration values, observations, and independently verified evidence without conflating them. Code and configuration reviews can identify changes to references, but they cannot prove undisclosed model internals or every execution path of an external API.
- For each release target and environment, identify the owners and information sources for models, data, and tools.
- Record the identifiers and timestamps you can obtain, and review differences from the previous record and unverified items.
- Depending on the change, check evaluations, permissions, and whether changes are reflected in search results, then pass the results and remaining risks to the approver.
Apply access controls to the inventory itself
An inventory contains information you may not want to disclose externally, such as connection targets and internal configurations. Do not indiscriminately record API keys, secrets, user prompts, or document bodies. Limit records to information necessary for the purpose, and define viewing permissions, retention periods, and deletion methods. Hashing does not necessarily make information anonymous. OWASP lists access controls and restrictions on data sources as measures against sensitive information exposure.
Common questions
If the model name stays the same, does that mean the AI BOM does not need updating?
No. Aliases or updates on the management side can change the actual version. Check the update policy for your environment and the version you can verify.
How should I record an external model’s training data when it is unknown?
Distinguish the provider’s public statements from independently verified facts, and record the information as undisclosed or unverified. Do not fill in gaps with guesses.
How Hi AUDIT Agent can help
When changing models or connection endpoints, it is also worth checking the security of the modified application code alongside updating the inventory. With Hi AUDIT Agent, you can request an analysis of the relevant code through chat in an MCP-compatible development environment and review the reasoning behind flagged issues and suggested fixes right there. You can request another analysis after making corrections, making it easier to incorporate changes into day-to-day reviews. Hi AUDIT Agent updates its detection capabilities and CVE data on the server, reducing the effort of updating its detection rules locally. Verify the model versions actually used and the production configuration separately, using operational records or other sources.
Create a Hi AUDIT Agent account