Ask the vendor which model or models power the feature, where they run, what customer data they receive, whether that data is retained or used for training, and how model changes are tested before release. Get the answers in writing, because they may change after you sign.
The six questions
- Which model, and whose? Is it a model the vendor built, a third party model accessed over an API, or a mix? You do not need a brand name in every case, but you need to know whether another company sees your data.
- Where does it run? In your environment, the vendor's cloud, or a third party's? Which regions?
- What does it see? Source code, dependency lists, findings only, or tickets and chat? Narrower is safer.
- Is customer data retained or used for training? Ask separately for prompts, outputs, and logs.
- What happens when the model changes? Providers update or retire models. Ask how the vendor tests a swap and whether you are told.
- What is the fallback? If the model service is down or unsure, does the product degrade gracefully?
Put the answers in a table
| Question | Acceptable | Needs follow up |
|---|---|---|
| Data sent | Minimal, documented fields | "Whatever is needed" |
| Retention | Stated period, deletion on request | Silence |
| Training use | Not used without opt in | Unclear or buried in terms |
| Model change | Tested and announced | Unannounced |
Why the model change question matters
An AI feature you tested in October may behave differently in March if the underlying model changed. That is not a defect, it is how hosted models work. The NIST AI Risk Management Framework, released in January 2023 for voluntary use, frames this as an ongoing job: the Measure and Manage functions are about tracking and responding to risk over time, not only at purchase.
Related data security guidance
In May 2025 CISA, with the NSA, the FBI, and international partners, published joint guidance on securing data used to train and operate AI systems, covering risks across the AI lifecycle. It is a useful reference when you need to explain to your own leadership why data handling questions belong in the evaluation.
What to do with vague answers
Do not fill the gaps with optimism. Mark the item as unanswered in your scorecard and ask your legal and privacy teams how to handle it. This post is not legal advice, and contract terms are a matter for your counsel.
Where Safeguard fits
Safeguard documents what its generally available features do, including SBOM and dependency scanning and reachability analysis, and offers a free scan at gold.safeguard.sh/scan. Ask us the same six questions and judge the answers.
Quick check
Do I need the exact model name?
You need to know who processes your data. The name helps but the data path matters more.
Is a vendor using a third party model a problem?
Not by itself. It adds a party to your data flow, so ask about it.
Sources
- NIST, AI Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework
- CISA, AI Data Security: Best Practices for Securing Data Used to Train and Operate AI Systems, published 22 May 2025: https://www.cisa.gov/resources-tools/resources/ai-data-security-best-practices-securing-data-used-train-operate-ai-systems
Last reviewed: 19 October 2026.
