How do we…
How do we check an AI vendor or model provider before we rely on it?
Add to board memoOpen the memo
The short answer
- Recommendation
- Ask the provider the AI security questions in writing, with evidence for every answer, and count the answers by what you hold. Check the contract carries the AI supplier terms, then give the counts to your supplier decision and ask again whenever the provider changes a model, a sub-processor or a term.
- Reasoning
- A provider's reply says what it does today, and your feature's security rests partly on it: where your data goes, who can read it, which model runs underneath and how you hear of a change. An answer is a claim until you have seen the evidence, so a question answered without evidence stays open. Counting answers with evidence, without evidence and not answered gives a result you can act on. The questionnaire informs the supplier decision under your third-party process and never replaces it.
- Considerations
- The only good answer on training is a plain no, written into the contract (ST-1). Without it the tool takes no Internal or Confidential data. A vendor setting verified off is an added control, checked at each review, and never a substitute for the term.
- A certificate that covers other parts of the company does not answer for the AI feature you use.
- The review procedure requires the questionnaire from Depth 2 when an outside provider supplies the model. At Depth 1 the provider's published security statement is read instead.
- For a model you download, check its publisher, source and file hash, and load it in a form that does not run code (GS-15).
The decision, step by step
Select any step for the reasoning behind it and where the pack covers it.
- Start
- Step
- Decision
- Outcome
- Loop back
Keyboard: Tab to a step, arrow keys move along the flow, Enter pins its explanation, + and − zoom, 0 fits the diagram, Esc closes. The text version below has the same content.
Where a step quotes a number, it is the default in the CISO Times pack. Your own standard may set a different one.
Read as text
The same flow written out: each step, its reasoning and where the pack covers it. 19 steps.
Start
We want to rely on an AI vendor or model provider
A provider's reply says what it does today, and the security of your feature rests partly on it: where your data goes, who can read it, which model runs underneath and how you hear of a change. The questionnaire helps you ask for these facts in a form you can count. The decision to use the supplier stays with your third-party process.
In the pack
- P25-A06 AI Vendor & Model Provider Security QuestionnaireInstructions
- P25-A03 LLM Application Security Review ProcedureStage 4 (RP-4): Assess the provider
Decision
A provider's service, or a model we download ourselves?
A hosted service or a tool with an AI feature has a provider who can answer the questions. A model or adapter you take from a public source has a publisher rather than a provider you contract with. If you do both, run each route. If you do not know which applies, take the provider route and ask who supplies the model.
- A provider's service or tool: Will it get Internal or Confidential data, or is the use Tier 2 or 3?
- A model or adapter we download: Check publisher, source and file hash, and load it safely
In the pack
- P25-A02 Generative AI Security StandardModel and provider (rule GS-15)
Outcome
Check publisher, source and file hash, and load it safely
Use a downloaded model or adapter only after you have checked who published it, where it came from and its file hash, and load it in a form that does not run code when loaded. Record the model name and version, and add it to the list of models, datasets, libraries and connectors with versions. Pin tool and connector servers and approve each one before use.
In the pack
- P25-A02 Generative AI Security StandardModel and provider (rules GS-15, GS-16)
Decision
Will it get Internal or Confidential data, or is the use Tier 2 or 3?
P24 asks for a supplier assessment before the supplier receives Internal or Confidential data, and before a Tier 2 or Tier 3 use starts even with Public data. If you do not know the data level or the tier, answer yes. A Tier 1 use with Public data only is the one case P24 leaves out.
- Yes, or not known: Name the provider and the company behind the model
- No: P24 asks for no supplier assessment yet
In the pack
- P24-A02 AI Acceptable Use & Security PolicyThe inventory and suppliers (rule AG-11)
- P24-A07 AI Tool Approval & Data Permission MatrixData Rules (rules BC-3, BC-4)
Outcome
P24 asks for no supplier assessment yet
P24 does not ask for a supplier assessment for a Tier 1 use that takes Public data only. Your own third-party process may still ask for one. A new purpose, a higher data level, a new audience or a change of supplier, model or location forces a review of the use, and the answer to this question is then asked again.
In the pack
- P24-A02 AI Acceptable Use & Security PolicyThe inventory and suppliers (rules AG-10, AG-11)
- P24-A07 AI Tool Approval & Data Permission MatrixData Rules
Step
Name the provider and the company behind the model
Use one workbook for each provider. If the company that supplies the model differs from the provider you contract with, name both and put the model questions to the provider about each. A tool supplier may run another party's model, and that party may keep or use your data on its own terms.
In the pack
- P25-A06 AI Vendor & Model Provider Security QuestionnaireProvider
- P25-A06 AI Vendor & Model Provider Security QuestionnaireQuestionnaire (rules VQ-04, VQ-13, VQ-18)
Step
Send the questions and ask for evidence with each answer
Send the question and evidence columns only, and keep the reviewer's good-answer columns back. Set a reply date. A trust centre often answers many rows: paste its link as the evidence and check it. Questions about agents, tools and connectors apply only when the feature has tools, and a question that does not apply needs a stated reason. The review procedure requires the questionnaire from Depth 2 when an outside provider supplies the model; at Depth 1 it asks only that the provider's published security statement is read.
In the pack
- P25-A06 AI Vendor & Model Provider Security QuestionnaireInstructions
- P06-A03 Supplier Security Assessment ProcedureStage 3 — Questionnaire
- P25-A06 AI Vendor & Model Provider Security QuestionnaireQuestionnaire (rules VQ-29, VQ-30, VQ-31)
Decision
Did the provider reply by the date you set?
Judgement: a late or partial reply is still a reply, and a blank row counts as not answered. The third-party pack escalates a supplier that sends no answer or evidence by the deadline, so this flow takes the no branch only for a supplier that sends nothing by then.
- Yes, in full or in part: Count the answers, then read each against a good answer
- No: No reply: escalate to the business owner first
In the pack
- P25-A06 AI Vendor & Model Provider Security QuestionnaireResult
- P06-A03 Supplier Security Assessment ProcedureStage 3 — Questionnaire
- P06-A03 Supplier Security Assessment ProcedureEscalation (rules TP-03, TP-04)
Outcome
No reply: escalate to the business owner first
The third-party pack escalates a supplier that does not answer or send evidence by the deadline to the business owner, and then to the Head of Information Security. Until the supplier replies, every question counts as not answered. Judgement: keep the use to the data levels the tool is already approved for until it does.
In the pack
- P06-A03 Supplier Security Assessment ProcedureEscalation (rules TP-03, TP-04)
- P24-A07 AI Tool Approval & Data Permission MatrixData Rules
Step
Count the answers, then read each against a good answer
Mark each row Answered with evidence, Answered without evidence, Not answered or Not applicable. The reviewer, never the feature's owner, reads each answer against the good answer and records Meets or Falls short; the review procedure has the Head of Information Security read them. A certificate that covers other parts of the company does not answer for the AI feature you use. In the distributor's example, 24 of 31 applicable questions were answered with evidence and 7 stayed open.
In the pack
- P25-A06 AI Vendor & Model Provider Security QuestionnaireResult
- P25-A06 AI Vendor & Model Provider Security QuestionnaireQuestionnaire (rules VQ-27, VQ-28)
- P25-A06 AI Vendor & Model Provider Security QuestionnaireDefinitions
- P25-A03 LLM Application Security Review ProcedureStage 4 (RP-4): Assess the provider (rule Head of Information Security reads)
Decision
Can our prompts or files be used to train any model?
This covers the provider and the company behind its model (VQ-01, VQ-02, VQ-04). The only good answer is a plain no, written into the contract. A product page or a setting, even one verified off, does not replace the contract term. If the answer is yes, not known or not in the contract, answer on that branch.
- No, barred in the contract: Does the contract carry the other AI terms, ST-2 to ST-8?
- Yes, not known, or not in the contract: Keep to Public data until ST-1 is in the contract
In the pack
- P25-A06 AI Vendor & Model Provider Security QuestionnaireQuestionnaire (rules VQ-01, VQ-02, VQ-04)
- P24-A07 AI Tool Approval & Data Permission MatrixData Rules (rules BC-2, ST-1)
Step
Keep to Public data until ST-1 is in the contract
Without a contract term that bars training (ST-1), the tool takes no Internal or Confidential data, only Public information. A vendor setting verified off does not replace the term: where the tool has one, keep it off and check it at each review as an added control. P06 has no clause for ST-1, so add it in your own words. A use already running gets a dated condition to add the term (H-2).
In the pack
- P24-A07 AI Tool Approval & Data Permission MatrixData Rules (rules BC-2, ST-1)
- P24-A02 AI Acceptable Use & Security PolicyWhat the contract must say
- P24-A03 AI Use Case Intake & Approval ProcedureHolds (rule H-2)
Decision
Does the contract carry the other AI terms, ST-2 to ST-8?
The terms cover processor terms, sub-processors, locations, incident notice, return and deletion on exit, notice of AI feature changes and the intended purpose. Each has a P06 clause except the statement of intended purpose (ST-8). Check the contract itself as well as the answers, because the standard asks for both. If you do not know what the contract says, answer no.
- Yes: Is any question open, or does any answer fall short?
- No, or not known: Take each gap to the contract by term and clause
In the pack
- P25-A06 AI Vendor & Model Provider Security QuestionnaireTerms
- P25-A02 Generative AI Security StandardModel and provider (rule GS-17)
- P06-A07 Supplier Contract Security Clause LibraryCL-04 Sub-processors and subcontracting
Step
Take each gap to the contract by term and clause
Use the Terms sheet to see which questions inform each term and clause. For example VQ-24 and VQ-26 inform ST-5 and CL-07, VQ-07, VQ-08 and VQ-32 inform ST-6 and CL-13, and VQ-17 informs ST-7 and CL-20. ST-8 has no P06 clause: add it in your own words. Procurement raises each gap with the supplier. A supplier that refuses a required clause goes to the exception route.
In the pack
- P25-A06 AI Vendor & Model Provider Security QuestionnaireTerms
- P06-A07 Supplier Contract Security Clause LibraryCL-07 Security incident notification
- P06-A07 Supplier Contract Security Clause LibraryCL-13 Data return and deletion
- P06-A07 Supplier Contract Security Clause LibraryCL-20 Notice of material changes
- P06-A03 Supplier Security Assessment ProcedureEscalation (rule TP-07)
Decision
Is any question open, or does any answer fall short?
Open means answered without evidence or not answered. Falls short means the reviewer judged the answer or its evidence weaker than the good answer. Judgement: read the counts by section as well as in total, because a gap in one section, such as the model supply chain, can sit behind good answers elsewhere. Never turn the counts into a percentage or a score.
- Yes: Agree a date with the provider to close the open questions
- No: Give the counts and answers to the supplier decision
In the pack
- P25-A06 AI Vendor & Model Provider Security QuestionnaireResult
- P25-A06 AI Vendor & Model Provider Security QuestionnaireDefinitions
Step
Agree a date with the provider to close the open questions
For each open question ask for the evidence named in the Evidence column, or ask again using the reviewer's ask-more column. Write the agreed date on the Provider sheet. An answer that falls short is a gap to take to the contract or to the supplier decision as a condition, not something to leave unread.
In the pack
- P25-A06 AI Vendor & Model Provider Security QuestionnaireInstructions
- P25-A06 AI Vendor & Model Provider Security QuestionnaireProvider
Step
Give the counts and answers to the supplier decision
The assessor writes the review report with the questionnaire and evidence, and the business owner decides and signs: approve, approve with conditions, escalate or reject. Open questions and shortfalls can become conditions with deadlines. The answers inform that decision and never replace it. Keep the completed workbook with the feature's review file.
In the pack
- P06-A03 Supplier Security Assessment ProcedureStage 7 — Decision
- P25-A03 LLM Application Security Review ProcedureStage 4 (RP-4): Assess the provider (rule E-8)
Decision
Has the provider since changed a model, sub-processor or term?
A provider's reply says what it does today. A change of model is a trigger to review the feature, and a supplier that adds or changes an AI feature gives notice under ST-7. If you learn of a change late, answer yes and ask again for the rows it touches.
- Yes: ask again: Send the questions and ask for evidence with each answer (back to an earlier step)
- No: Reviewed: keep the workbook and watch for changes
In the pack
- P25-A06 AI Vendor & Model Provider Security QuestionnaireInstructions
- P25-A02 Generative AI Security StandardChange control (rule GS-18)
Outcome
Reviewed: keep the workbook and watch for changes
Keep the completed workbook with the review evidence, as the standard asks. Watch for a notice of a new model, a new sub-processor or a changed term. Each is a reason to ask again and to review the feature within the working days the standard sets.
In the pack
- P25-A02 Generative AI Security StandardEvidence (rule GS-29)
- P25-A03 LLM Application Security Review ProcedureWhat forces a new review
Practical checklist
- Name the provider you contract with and the company behind its model.
- Send the questions with a reply date, and ask for evidence with each answer.
- Escalate a supplier that sends nothing by the date to the business owner, then the Head of Information Security.
- Have someone other than the feature's owner count each answer and read it against a good answer.
- Confirm the contract bars training on your prompts and files, or keep the tool to Public information until it does.
- Check the contract itself for the other AI terms, and raise each gap through Procurement.
- Agree a date with the provider to close each open question.
- Give the counts and answers to the supplier decision, and ask again after any change.
Evidence to have ready
- The completed questionnaire with the evidence for each answer and the reviewer's reading.
- The counts by section: answered with evidence, answered without evidence, not answered.
- The contract showing the AI terms and the supplier clauses, with each gap raised.
- The agreed dates for open questions, and the signed supplier decision with any conditions.
Assumptions
- You contract with a provider for a hosted AI service, or for software with an AI feature, and the use is on the AI inventory or going through intake.
- Your third-party process already tiers and assesses suppliers and has its own security questionnaire. Judgement: this flow adds the AI-specific questions to that work and does not replace it.
- The people reading the answers are not the owner of the feature. An answer given without evidence stays open, even when the provider is the only party who can answer.
- The terms ST-1 to ST-8 and the clause IDs are the packs' defaults. Your own contract standard may ask for more.
Not covered here
- What to require when a vendor may train on your data. That is “If an AI vendor may train on our data, then what do we require?”.
- Tiering and assessing suppliers in general. Your third-party risk process covers it.
- Security review of an LLM feature you build yourself. That is “How do we assess the security of a GenAI feature we are building ourselves?”.
- Whether the AI use may go ahead at all. That is “How do we decide whether a proposed AI use case can go ahead?”.
Related questions
- If an AI vendor may train on our data, then what do we require?
- How do we decide whether a proposed AI use case can go ahead?
- How do we assess the security of a GenAI feature we are building ourselves?
- How do we review an LLM feature or an AI agent for security before it goes live?
- How do we track AI security findings and the deviations we accept?