When a vendor with access to your systems or nonpublic information uses AI, NYDFS guidance says to consider the AI threats it faces, how they could affect you, and how it protects itself. Ask what data reaches models, who can access it, how incidents are notified, and what the contract warrants.
This is general information, not legal advice.
Where the questions come from
Section 500.11 of Part 500 requires third party service provider policies covering due diligence, minimum practices and contract protections on access controls, encryption, incident notice and representations. The October 2024 industry letter adds that diligence should consider AI related threats, and that if providers use AI you should consider extra representations and warranties about secure use of your nonpublic information, including taking advantage of available enhanced privacy, security and confidentiality options.
Eight questions
| # | Question | What a good answer includes |
|---|---|---|
| 1 | Which features use AI, and what do they do? | A specific list, not a general claim |
| 2 | What of our data reaches a model or a model provider? | Fields, volume, and whether data leaves the vendor's boundary |
| 3 | Is our data used to train or improve models? | A clear yes or no, backed by contract text |
| 4 | How long is it retained? | A period, and a deletion process |
| 5 | Who at the vendor can see prompts and outputs? | Role based access and logging |
| 6 | How is access protected? | MFA and encryption statements aligned to sections 500.12 and 500.15 |
| 7 | How do you test the AI feature for security flaws? | A testing process and owner |
| 8 | When will you tell us about an event? | A notice period and a named contact |
Reading the answers
Specific answers with documents behind them carry weight. General statements such as "we take privacy seriously" do not tell you what happens to your data. Ask whether the answers can be written into the agreement as representations.
Use the enhanced options
The letter points to enhanced privacy, security and confidentiality options that providers may offer. If a vendor has a setting that keeps your data out of model training or limits retention, ask for it to be enabled and recorded in the contract or order form.
Reassess when AI is added later
Many vendors add AI features after you sign. Include a clause that requires notice of material new AI use, and re-run the questions when it arrives. That fits the periodic reassessment in section 500.11(a)(4).
Mistakes to avoid
Reviews often accept a one line statement about training on customer data without a contract clause behind it. They miss that AI features can run through another provider, so the real data path includes a fourth party. And they treat the answer as permanent. Ask who processes the data downstream, require the answer in writing in the agreement, and set a trigger for review whenever a vendor announces a new AI feature.
Quick check
Is there a required AI questionnaire?
No. The guidance recommends considering AI threats in diligence and contract terms. The questions are yours to design.
Who should own these answers?
Vendor management usually owns the file, with security and engineering supplying technical review.
Sources
- NYDFS Industry Letter on AI cybersecurity risks, October 16, 2024
- 23 NYCRR 500.11, Third-party service provider security policy
Last reviewed: 19 October 2026.
