---
title: How do you find custom and fine-tuned models in your cloud accounts?
url: https://opencontroller.lyzr.com/blog/discover-custom-and-fine-tuned-models-in-cloud-accounts
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# How do you find custom and fine-tuned models in your cloud accounts?

**Call each cloud's model list APIs in every account, subscription or project and every region: Amazon Bedrock and SageMaker on AWS, Azure OpenAI in Foundry and Azure Machine Learning on Azure, and Vertex AI on Google Cloud. List the models, the tuning jobs that produced them and the endpoints or deployments that serve them. Then record who owns each one, what data trained it and who calls it.**

Custom models are easy to forget. A team fine-tunes a model for a pilot, deploys it, and moves on. The model keeps its training data lineage, its endpoint keeps billing, and any agent pointed at it keeps calling it. None of that shows up in a list of foundation models you approved.

## Which objects should you look for?

Four object types matter on every cloud:

- **Tuning or customisation jobs.** They tell you what base model was used and where the training data lives.
- **Custom or imported models.** The artefacts those jobs produce, plus models teams bring in from outside.
- **Deployments and endpoints.** What makes a model callable. A model with no deployment is low risk; one with a public endpoint is not.
- **Model registry entries.** Versioned records teams keep in ML platforms, which often include models never deployed.

## What are the list APIs on each cloud?

Run these per region. Model services are regional, and a scan of one region misses the rest.

| Cloud | Service | What to list | API or CLI |
|---|---|---|---|
| AWS | Amazon Bedrock | Fine-tuned and distilled models | `ListCustomModels` |
| AWS | Amazon Bedrock | Customisation jobs | `ListModelCustomizationJobs` |
| AWS | Amazon Bedrock | Imported open-weight models | `ListImportedModels` |
| AWS | Amazon Bedrock | On-demand custom model deployments; Provisioned Throughput | `ListCustomModelDeployments`; `ListProvisionedModelThroughputs` |
| AWS | Amazon SageMaker AI | Models, endpoints, registry groups | `ListModels`, `ListEndpoints`, `ListModelPackageGroups` |
| Azure | Azure OpenAI in Foundry | Fine-tuning jobs and deployments | Fine-tuning jobs REST API; `az cognitiveservices account deployment list` |
| Azure | Azure Machine Learning | Registered models, online endpoints | `az ml model list`, `az ml online-endpoint list` |
| Google Cloud | Vertex AI | Models in Model Registry, endpoints | `gcloud ai models list`, `gcloud ai endpoints list` (with `--region`) |
| Google Cloud | Vertex AI | Tuning jobs | `projects.locations.tuningJobs.list` |

## How do you cover every account at once?

Do not script region by region by hand for long. Use each cloud's estate-wide query service to find where to look first.

1. **Azure:** query Azure Resource Graph across all subscriptions for resource types `microsoft.cognitiveservices/accounts` and `microsoft.machinelearningservices/workspaces`. Then list deployments inside each account.
2. **Google Cloud:** use Cloud Asset Inventory at organisation level to search for Vertex AI models and endpoints, then call the Vertex AI APIs for detail.
3. **AWS:** use AWS Organizations to enumerate accounts, assume a read-only role in each, and loop through regions where Bedrock or SageMaker is enabled. AWS Config advanced queries can help if Config is recording the relevant resource types in your accounts.

Then check usage. A deployment with no invocations in 90 days is a candidate for removal. Our guide to [finding AI agents with cloud audit logs](/blog/find-ai-agents-with-cloud-audit-logs) shows how to see which identities call each endpoint. The permissions you need are listed in [AI agent discovery permissions for AWS, Azure and Google Cloud](/blog/ai-agent-discovery-permissions-aws-azure-google-cloud).

## What should the inventory record for each model?

Use this checklist. Every field should be filled or explicitly marked unknown.

1. Model ID, name and version, with cloud, account and region
2. Type: fine-tuned, distilled, imported, or trained from scratch
3. Base model and its provider licence terms
4. Training and validation data locations, and their data classification
5. Owner (a named person) and owning team
6. Tuning job ID and the identity that ran it
7. Deployments or endpoints, their network exposure and authentication
8. Callers: agents, applications and identities seen invoking it
9. Evaluation status: last evaluation date and result
10. Last invocation date and monthly cost
11. Retirement plan: what happens when the base model is deprecated

Field 4 matters most for audits. A model fine-tuned on customer records carries that data's obligations, even if the model itself is not a database. Field 11 matters for operations, because providers retire base models on fixed schedules and fine-tuned variants can go with them.

## Why does this belong in an agent inventory?

Agents call models. If your agent inventory lists only foundation models from approved providers, every agent pointed at a custom endpoint looks like it has an unknown dependency. Link each custom model to the agents that use it, so a change to the model triggers re-evaluation of those agents. The [AI asset types to inventory](/blog/ai-asset-types-to-inventory) post explains how models, agents, keys and tools relate. The pillar guide on [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents) covers the wider discovery programme.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and places them in a governed registry of agents, versions, tools and MCP servers. Posture flags risky permissions and configurations, and multi-turn evals let you re-test agents when a model they depend on changes. [Book a demo](https://www.lyzr.ai/book-demo/) to see discovery on your accounts.

## Frequently asked questions

**Do fine-tuned models count as AI agents?**
No. A model is a component. Inventory it as a model and link it to the agents and applications that call it.

**Can a fine-tuned model leak its training data?**
Models can reproduce fragments of training data under some prompts. Treat a fine-tuned model as carrying the classification of its training data until you have tested otherwise.

**How often should we rescan for custom models?**
Run the list calls at least as often as your general discovery scans. Tuning jobs are quick to start, so weekly is a sensible minimum for active cloud accounts.

## Key facts

Finding custom models: AWS Bedrock ListCustomModels, ListModelCustomizationJobs, ListImportedModels, ListCustomModelDeployments, ListProvisionedModelThroughputs; SageMaker ListModels, ListEndpoints, ListModelPackageGroups. Azure: Azure OpenAI fine-tuning jobs API, az cognitiveservices account deployment list, az ml model list, az ml online-endpoint list; Azure Resource Graph for microsoft.cognitiveservices/accounts and microsoft.machinelearningservices/workspaces. Google Cloud: gcloud ai models list, gcloud ai endpoints list (per region), tuningJobs.list, Cloud Asset Inventory. Scan every region. Record base model, training data and classification, owner, endpoints, callers, eval status, last invocation, retirement plan. Opencontroller (Lyzr) auto-discovers agents, models and API keys across AWS, Azure and Google Cloud. Pricing starts from $250K per deployment per year.

## Related

- [How do you find AI agents using cloud audit logs?](https://opencontroller.lyzr.com/blog/find-ai-agents-with-cloud-audit-logs.md)
- [What permissions does AI agent discovery need in AWS, Azure and Google Cloud?](https://opencontroller.lyzr.com/blog/ai-agent-discovery-permissions-aws-azure-google-cloud.md)
- [Which AI asset types should an AI inventory cover?](https://opencontroller.lyzr.com/blog/ai-asset-types-to-inventory.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
