AI vendor risk assessment questionnaire

1. DU-1Do you use our data (inputs, outputs, files, feedback or metadata) to train, fine-tune or improve any model, yours or a third party’s? Is that off by default, and where does the contract say so?

   Training on your data can expose confidential information to other customers and creates IP and privacy risk. NIST says third-party policies should cover these risks, and the banking agencies list confidentiality and ownership as contract terms.

   Red flags: Opt-out only in a settings page, not the contract. Undefined "aggregated" or "de-identified" use. Terms that let the vendor change data use by notice.

   Asked of every AI vendor. Sources: NIST AI RMF GOVERN 6.1; MAP 4.1; Interagency TPRM guidance (2023) Contract: confidentiality and integrity; ownership and license

2. DU-2If our data has been or could be used for training, can we opt out in the contract, and can you confirm in writing whether our data is in any model already trained?

   Once data is in a trained model it is hard to remove. You need a contractual opt-out and a clear record of past use.

   Red flags: Cannot say whether past data was used. No way to stop future use without leaving the service.

   Asked when: training on your data. Sources: NIST AI RMF MAP 4.1; MANAGE 3.1; ISO/IEC 42001 A.7.5 Data provenance; GDPR Art. 28 Art. 28(3)(a)

3. DU-3Where are our data, prompts and outputs stored and processed (countries and regions), and how long are they kept, including in logs, backups and abuse-monitoring stores?

   Retention and location drive breach exposure, cross-border transfer duties and what you can delete on request.

   Red flags: Different retention for "safety" logs that is not disclosed. Processing in regions not named in the contract.

   Asked of every AI vendor. Sources: Interagency TPRM guidance (2023) Contract: responsibilities for providing, receiving and retaining information; foreign-based third parties; GDPR Art. 28 Art. 28(3)(a) (instructions incl. transfers)

4. DU-4Who owns the outputs, and what licence do you take over our inputs and outputs?

   Broad licences can let a vendor reuse your content. Ownership of outputs matters if you build products on them.

   Red flags: A perpetual, worldwide licence to use customer content "to provide and improve services". No statement on output ownership.

   Asked of every AI vendor. Sources: Interagency TPRM guidance (2023) Contract: ownership and license; NIST AI RMF GOVERN 6.1 (IP risk)

5. DU-5Will you process personal data only on our documented instructions, as our processor or service provider, and sign a data processing agreement?

   Under GDPR a processor must act only on the controller’s documented instructions, including for transfers outside the EU, and the terms must be in a binding contract.

   Red flags: Refuses a DPA, or claims to be an independent controller of your users’ data.

   Asked when: personal data. Sources: GDPR Art. 28 Art. 28(1), 28(3); Interagency TPRM guidance (2023) Due diligence: legal and regulatory compliance

6. DU-6Will you sign a HIPAA business associate agreement that covers this product and every AI feature we will use?

   A covered entity may disclose PHI to a business associate only with a contract that meets 45 CFR 164.504(e). AI features outside the BAA cannot receive PHI.

   Red flags: BAA excludes AI, beta or third-party model features. BAA only on a tier you are not buying.

   Asked when: PHI or HIPAA. Sources: HIPAA (45 CFR 164.502(e), 164.504(e)) 45 CFR 164.502(e), 164.504(e)

7. DU-7Is PHI excluded from model training and from review by your staff except as the BAA permits, and will you return or destroy it when the contract ends?

   A business associate may use PHI only as the contract permits and must return or destroy it at termination where feasible.

   Red flags: PHI used for "product improvement". No deletion commitment for model inputs and logs.

   Asked when: PHI. Sources: HIPAA (45 CFR 164.502(e), 164.504(e)) Permitted uses; return or destruction at termination

1. SE-1Which independent security reports or certifications cover this product (for example SOC 2 Type 2, ISO/IEC 27001, ISO/IEC 42001), and does their scope include the AI features and the systems that run them?

   An attestation only covers what is in scope. AI features are often run by subprocessors outside the audited system.

   Red flags: Report scope excludes the AI feature or its model provider. Certificate held by a different legal entity. Report more than 12 months old.

   Asked of every AI vendor. Sources: Interagency TPRM guidance (2023) Due diligence: information security; risk management; NIST AI RMF MAP 4.2

2. SE-2Do you support SSO, MFA, role-based access and audit logs of user and admin activity that we can export?

   You need to control who can use the AI and reconstruct what happened after an incident.

   Red flags: Shared API keys only. Audit logs not available to customers or kept for days.

   Asked of every AI vendor. Sources: Interagency TPRM guidance (2023) Due diligence: information security; management of information systems

3. SE-3How do you defend against prompt injection, including indirect injection from documents, emails or web pages the model reads? What testing evidence can you share?

   Prompt injection is first on OWASP’s 2025 list of LLM application risks. It can make the model ignore instructions or leak data.

   Red flags: Relies only on the system prompt. No testing evidence. Treats retrieved content as trusted.

   Asked when: generative AI or agents. Sources: OWASP Top 10 for LLM Apps 2025 LLM01:2025 Prompt Injection; NIST AI RMF MEASURE 2.7

4. SE-4How do you stop our data, other customers’ data or secrets from leaking through model outputs, logs or the system prompt?

   OWASP lists sensitive information disclosure and system prompt leakage among the top LLM risks.

   Red flags: Customer data shares a model context or cache with other tenants. Secrets placed in prompts.

   Asked when: generative AI. Sources: OWASP Top 10 for LLM Apps 2025 LLM02:2025; LLM07:2025

5. SE-5Which tools, systems and permissions can the agent use, and can we restrict them to the minimum each task needs?

   Excessive agency, meaning too much functionality, permission or autonomy, is one of OWASP’s top LLM risks.

   Red flags: Agent runs with an admin or broad service account. No per-tool allowlist. Permissions cannot be scoped by the customer.

   Asked when: agents. Sources: OWASP Top 10 for LLM Apps 2025 LLM06:2025 Excessive Agency; NIST AI RMF MAP 4.2

6. SE-6How do you vet and monitor the third-party models, datasets, plugins and open-source components in the product?

   Supply-chain weaknesses are on OWASP’s list, and NIST says pre-trained models should be monitored as part of maintenance.

   Red flags: No inventory of model and component versions. Unvetted community models or plugins.

   Asked of every AI vendor. Sources: OWASP Top 10 for LLM Apps 2025 LLM03:2025 Supply Chain; NIST AI RMF MANAGE 3.2

1. MR-1Which models power the product (yours or a provider’s, with names and versions), and how and when will you tell us before they change?

   A model swap can change accuracy, bias and behaviour overnight. You cannot assess what you cannot name.

   Red flags: Will not name the underlying model. Model changes without notice. No way to pin a version.

   Asked of every AI vendor. Sources: NIST AI RMF MANAGE 3.2; Colorado SB 26-189 Developer notice of material updates; EU AI Act Art. 13(3)(c)

2. MR-2What testing do you do before each release (accuracy, robustness, red-teaming), and will you share results relevant to our use case?

   Performance should be shown under conditions like yours, not only on public benchmarks.

   Red flags: Only marketing benchmarks. No pass criteria set in advance. Results withheld entirely.

   Asked of every AI vendor. Sources: NIST AI RMF MEASURE 2.3; MEASURE 2.7; EU AI Act Art. 9(6)-(8); Art. 13(3)(b)(ii)

3. MR-3Have you tested for bias across groups relevant to our use, and will you share the method and results?

   Systems that influence decisions about people can produce unfair outcomes. You will need evidence for your own assessments.

   Red flags: "We don’t collect demographic data, so we can’t test." Results only for a different use case.

   Asked when: consequential decisions or EU high-risk use. Sources: NIST AI RMF MEASURE 2.11; EU AI Act Art. 13(3)(b)(v); Colorado SB 26-189 Developer documentation: known limitations and risks

4. MR-4How do you measure and reduce incorrect or ungrounded outputs (hallucinations) for our use case?

   OWASP lists misinformation from LLMs as a top risk, and NIST asks for evidence that a system is valid and reliable.

   Red flags: No measured error rate. Responsibility placed entirely on the user.

   Asked when: generative AI. Sources: OWASP Top 10 for LLM Apps 2025 LLM09:2025 Misinformation; NIST AI RMF MEASURE 2.5

5. MR-5Will you give us what we need to validate the model under our model risk program: design and theory, development data, performance testing and ongoing monitoring results?

   SR 26-2, which replaced SR 11-7 in April 2026, says firms should understand a vendor model’s design, data and performance and monitor it. It puts generative and agentic AI outside its scope, but expects broader risk management to cover them.

   Red flags: "Proprietary" with no substitute evidence such as an independent validation report.

   Asked when: financial services. Sources: SR 26-2 (replaces SR 11-7) Section VII (vendor products); Interagency TPRM guidance (2023) Due diligence: risk management

6. MR-6Do you monitor model performance and drift in production, and will you share monitoring results and alerts with us?

   NIST asks deployers to monitor the risks and benefits of third-party resources over time. Contract performance measures need data behind them.

   Red flags: No production monitoring. Metrics available only on request.

   Asked of every AI vendor. Sources: NIST AI RMF MANAGE 3.1; MEASURE 2.4; Interagency TPRM guidance (2023) Contract: performance measures or benchmarks

1. TD-1Will you document the product’s intended uses, known limitations, foreseeable misuse and residual risks?

   NIST says negative residual risks should be documented for downstream acquirers. Without this you cannot set safe use rules.

   Red flags: Only marketing material. "No known limitations."

   Asked of every AI vendor. Sources: NIST AI RMF MANAGE 1.4; ISO/IEC 42001 A.8.2 System documentation and information for users; EU AI Act Art. 13(3)(b)

2. TD-2Will you provide instructions for use that meet Article 13: accuracy metrics, robustness and cybersecurity levels, human oversight measures and logging mechanisms?

   Deployers of high-risk systems must use them according to the instructions for use, so the instructions must be complete. Annex III obligations apply from December 2, 2027.

   Red flags: No instructions for use, or instructions without accuracy metrics or oversight measures.

   Asked when: EU + EU high-risk use. Sources: EU AI Act Art. 13(3); Art. 26(1)

3. TD-3Does the product tell people they are interacting with AI where required, and mark AI-generated audio, image, video or text in a machine-readable way?

   Article 50 duties have applied since August 2, 2026. Generative systems already on the market before then have until December 2, 2026 to mark outputs.

   Red flags: No marking, or no plan to meet the December 2, 2026 deadline for existing systems.

   Asked when: EU, and generative AI. Sources: EU AI Act Art. 50(1)-(2)

4. TD-4Will you give us the developer documentation Colorado SB 26-189 requires: intended uses, known harmful uses, training-data categories, known limitations and risks, instructions for appropriate use and human review, and notice of material updates?

   Colorado deployers must notify consumers and explain adverse outcomes from January 1, 2027, and need this documentation to do it. Attorney General rules are not yet final.

   Red flags: Vendor says the law does not apply without explaining why. No commitment before January 1, 2027.

   Asked when: US state laws + consequential decisions. Sources: Colorado SB 26-189 Developer duties

5. TD-5What categories and sources of data were used to train or fine-tune the models, and on what licences?

   Training-data provenance affects IP exposure, bias and fitness for your population.

   Red flags: No information at all, even at category level. Scraped data with no licence position.

   Asked of every AI vendor. Sources: NIST AI RMF MAP 4.1; ISO/IEC 42001 A.7.5 Data provenance; Colorado SB 26-189 Training-data categories

1. HO-1What controls let our staff review, override or reverse outputs before they take effect?

   Human oversight only works if the product lets people intervene. For EU high-risk systems, overseers must be able to disregard, override or reverse outputs.

   Red flags: Outputs act automatically with no review step. Overrides not logged.

   Asked of every AI vendor. Sources: NIST AI RMF MAP 3.5; GOVERN 3.2; EU AI Act Art. 14(4)(d)

2. HO-2Which agent actions require human approval, and can we set approval thresholds, spending limits and a kill switch?

   NIST asks for ways to deactivate systems that perform outside their intended use. For EU high-risk systems, overseers must be able to stop the system safely.

   Red flags: No approval step for irreversible actions. No way to stop a running agent.

   Asked when: agents. Sources: NIST AI RMF MANAGE 2.4; OWASP Top 10 for LLM Apps 2025 LLM06:2025 Excessive Agency; EU AI Act Art. 14(4)(e)

3. HO-3What training and material will you provide so the people we assign to oversight have the competence the law requires?

   EU deployers must assign oversight to people with the necessary competence, training and authority, and take measures to support AI literacy.

   Red flags: No training material. Oversight guidance limited to "review outputs".

   Asked when: EU + EU high-risk use. Sources: EU AI Act Art. 26(2); Art. 4

4. HO-4Does the product let affected people get an explanation of a decision and request human review?

   Colorado deployers must explain adverse outcomes and offer human review where commercially reasonable. EU deployers of Annex III systems must tell people they are subject to the system.

   Red flags: No per-decision explanation. No workflow for review requests.

   Asked when: consequential decisions. Sources: NIST AI RMF MEASURE 3.3; Colorado SB 26-189 Deployer duties: adverse-outcome explanation; human review; EU AI Act Art. 26(11)

1. IN-1How quickly, and through what channel, will you notify us of security incidents, AI malfunctions or harmful outputs that affect us? Is the timeline in the contract?

   NIST asks for contingency processes for failures in third-party AI. The banking agencies list incident reporting and management in due diligence.

   Red flags: Notice only "as required by law". No named contact. No timeline for AI malfunctions as opposed to data breaches.

   Asked of every AI vendor. Sources: NIST AI RMF GOVERN 6.2; MANAGE 4.3; ISO/IEC 42001 A.8.4 Communication of incidents; Interagency TPRM guidance (2023) Due diligence: incident reporting and management processes

2. IN-2If a serious incident occurs, will you support our Article 26(5) duties and meet your own Article 73 reporting deadlines (15 days, 2 days for a widespread infringement, 10 days for a death)?

   Deployers who detect a serious incident must inform the provider immediately; providers report to market surveillance authorities within the Article 73 deadlines.

   Red flags: No incident contact for deployers. No process mapped to Article 73.

   Asked when: EU + EU high-risk use. Sources: EU AI Act Art. 26(5); Art. 73(2)-(4)

3. IN-3Will you report breaches of unsecured PHI and security incidents to us as 45 CFR 164.410 and the BAA require?

   Business associates must report breaches of unsecured PHI and security incidents to the covered entity.

   Red flags: Excludes "unsuccessful" incidents with no definition. Reporting only after the vendor’s own investigation ends.

   Asked when: PHI or HIPAA. Sources: HIPAA (45 CFR 164.502(e), 164.504(e)) 45 CFR 164.410; reporting provisions

4. IN-4Will you notify us of personal data breaches promptly and help us meet our GDPR breach-notification duties?

   Processors must assist controllers with Articles 32 to 36, which include breach notification.

   Red flags: No breach-notification timeline in the DPA.

   Asked when: EU + personal data. Sources: GDPR Art. 28 Art. 28(3)(f)

5. IN-5What happens if your model provider or service fails? Do you have a tested business continuity plan and a fallback model or mode?

   Many AI products depend on one model provider. The banking agencies list operational resilience and business continuity in due diligence and contracts.

   Red flags: Single model provider with no fallback. Continuity plan never tested.

   Asked of every AI vendor. Sources: Interagency TPRM guidance (2023) Due diligence: operational resilience; Contract: operational resilience and business continuity; NIST AI RMF GOVERN 6.2

1. SP-1List every subprocessor that touches our data, including model providers (which LLM APIs), hosting, monitoring and human reviewers.

   Your data goes wherever the vendor’s subprocessors go. The banking agencies list reliance on subcontractors in due diligence.

   Red flags: Model provider not named. List "available on request" only.

   Asked of every AI vendor. Sources: Interagency TPRM guidance (2023) Due diligence: reliance on subcontractors; NIST AI RMF GOVERN 6.1; ISO/IEC 42001 A.10.3 Suppliers

2. SP-2Will you notify us before adding or replacing a subprocessor, and can we object?

   Under GDPR a processor needs prior written authorisation for sub-processors and, under a general authorisation, must give notice of changes so the controller can object.

   Red flags: Changes posted to a web page with no notice. No right to object or terminate.

   Asked of every AI vendor. Sources: Interagency TPRM guidance (2023) Contract: subcontracting; GDPR Art. 28 Art. 28(2)

3. SP-3Do your model providers retain our data or use it for training? Do you have zero-retention or no-training terms with them?

   The vendor’s promises mean little if its own model provider keeps or trains on your data.

   Red flags: Vendor does not know its provider’s retention terms. Abuse-monitoring retention not disclosed.

   Asked when: generative AI. Sources: NIST AI RMF GOVERN 6.1; MANAGE 3.2; OWASP Top 10 for LLM Apps 2025 LLM03:2025 Supply Chain

4. SP-4Do subprocessors sign the same data protection terms you sign with us, and do you remain liable for them?

   GDPR requires the same obligations to flow down to sub-processors, with the processor liable for their failures. HIPAA requires subcontractors that handle PHI to agree to the same restrictions.

   Red flags: Weaker terms for subprocessors. Liability excluded for subprocessor failures.

   Asked when: personal data or PHI. Sources: Interagency TPRM guidance (2023) Contract: subcontracting; GDPR Art. 28 Art. 28(4); HIPAA (45 CFR 164.502(e), 164.504(e)) 45 CFR 164.502(e)(1)(ii)

1. CT-1Will the contract give us audit rights, or access to independent assurance reports, and require you to fix findings?

   The banking agencies list the right to audit and require remediation as a contract factor; GDPR requires processors to allow audits.

   Red flags: No audit right at all. Reports only under a separate paid agreement.

   Asked of every AI vendor. Sources: Interagency TPRM guidance (2023) Contract: right to audit and require remediation; GDPR Art. 28 Art. 28(3)(h)

2. CT-2Will you indemnify us for third-party IP claims arising from the model, its training data or its outputs? What is the liability cap?

   IP risk from AI outputs and training data is a named third-party risk in the NIST AI RMF.

   Red flags: Indemnity excludes AI outputs. Cap equal to a few months of fees.

   Asked of every AI vendor. Sources: Interagency TPRM guidance (2023) Contract: indemnification and limits on liability; NIST AI RMF GOVERN 6.1

3. CT-3At termination, how do we get our data back, in what format, and by when will you delete it, with written confirmation?

   Exit terms decide whether you can leave without losing data or leaving copies behind.

   Red flags: No export format. No deletion deadline. Copies kept in models or logs.

   Asked of every AI vendor. Sources: Interagency TPRM guidance (2023) Contract: default and termination; GDPR Art. 28 Art. 28(3)(g); HIPAA (45 CFR 164.502(e), 164.504(e)) Return or destruction of PHI

4. CT-4Are service levels and performance measures for the AI features in the contract (availability, latency, quality)?

   Performance measures are a named contract factor. Without them you have no remedy when quality drops.

   Red flags: SLA excludes AI features. "Best effort" only.

   Asked of every AI vendor. Sources: Interagency TPRM guidance (2023) Contract: performance measures or benchmarks

5. CT-5How does the contract split responsibility for complying with laws that apply to the AI system, between you, us and your suppliers?

   ISO/IEC 42001 has a control for allocating responsibilities with third parties. Under the EU AI Act, a provider of a high-risk system and its suppliers must agree in writing on the information and help the provider needs.

   Red flags: All compliance pushed to the customer. Silence on who does what.

   Asked of every AI vendor. Sources: Interagency TPRM guidance (2023) Contract: responsibility for compliance with applicable laws and regulations; ISO/IEC 42001 A.10.2 Allocating responsibilities; EU AI Act Art. 25(4)

6. CT-6Can you change the AI functionality, models or terms on your own? How much notice do we get, and can we terminate if a change is material?

   AI products change often. Unilateral changes can undo the assessment you just did.

   Red flags: Terms can change by posting online. No termination right for material changes.

   Asked of every AI vendor. Sources: Interagency TPRM guidance (2023) Contract: nature and scope of arrangement; NIST AI RMF MANAGE 3.1

7. CT-7Will the contract confirm that our regulators can examine the services you provide to us?

   The banking agencies list regulatory supervision as a contract factor.

   Red flags: Refuses to acknowledge regulator access.

   Asked when: financial services. Sources: Interagency TPRM guidance (2023) Contract: regulatory supervision

1. RG-1How do you classify the product under the EU AI Act (prohibited, high-risk, transparency duties, general-purpose AI), and what is your role?

   Your duties depend on the classification. If you rebrand the system, substantially modify it or change its purpose to a high-risk one, you can become its provider.

   Red flags: No classification done. Classification that ignores how you plan to use it.

   Asked when: EU. Sources: EU AI Act Art. 25(1); Art. 26

2. RG-2What material do you provide to help our staff understand the product’s capabilities and risks?

   Deployers must take measures to support the AI literacy of staff using AI. Since the Digital Omnibus, that does not require guaranteeing a specific level.

   Red flags: No user documentation or training beyond a product tour.

   Asked when: EU. Sources: EU AI Act Art. 4 (as amended by Reg. (EU) 2026/1744)

3. RG-3What is your plan for conformity assessment and EU database registration before the Annex III obligations apply on December 2, 2027?

   Public bodies must not use a high-risk system that is not registered in the EU database. Other deployers need assurance the provider will meet its duties.

   Red flags: No plan or timeline. Unaware of the December 2, 2027 date.

   Asked when: EU + EU high-risk use. Sources: EU AI Act Art. 26(8); Art. 49

4. RG-4Will you give us the information we need for a fundamental rights impact assessment and a data protection impact assessment?

   Some deployers must do a fundamental rights impact assessment before use, and deployers use the provider’s Article 13 information for DPIAs.

   Red flags: Refuses to share information on affected groups, risks or oversight.

   Asked when: EU + EU high-risk use. Sources: EU AI Act Art. 27(1); Art. 26(9)

5. RG-5Which US state AI laws do you think apply to the product, and how do you support deployers’ obligations under them?

   Colorado’s SB 26-189 takes effect January 1, 2027, and other states regulate AI in hiring, insurance and other areas.

   Red flags: "No state laws apply" with no analysis. No plan for Colorado.

   Asked when: US state laws. Sources: Colorado SB 26-189 Developer and deployer duties

Due diligence on a supplier whose product uses AI, before you buy and while you use it. On top of the usual security and privacy checks, it asks how the vendor uses your data for training, which models sit underneath, how they are tested and changed, how people can oversee and override outputs, and how incidents are reported. The NIST AI RMF asks organizations to have policies for third-party AI risks (GOVERN 6.1) and to monitor third-party resources over time (MANAGE 3.1).

Security questionnaires focus on controls such as access, encryption and incident response. AI products add risks those questionnaires rarely cover: training on customer data, silent model changes, prompt injection, hallucinated outputs, agents taking actions, and fourth-party model providers. Use this alongside your existing security questionnaire rather than instead of it.

The 22 core questions this tool always includes. The most important: whether the vendor trains on your data and where the contract says so; which models power the product and how you will hear about changes; every subprocessor, including model providers; how fast they will notify you of incidents; and what happens to your data when the contract ends.

Deployers of high-risk systems must use them according to the provider’s instructions for use, assign trained human oversight, monitor operation, keep logs they control for at least six months and report serious incidents to the provider (Article 26). Some deployers must also do a fundamental rights impact assessment (Article 27). If you rebrand a high-risk system, substantially modify it or change an AI system’s purpose so it becomes high-risk, you can become its provider (Article 25(1)). Annex III high-risk obligations apply from December 2, 2027.

No. On April 17, 2026 the Federal Reserve, OCC and FDIC replaced SR 11-7 with SR 26-2, Revised Guidance on Model Risk Management. SR 26-2 still says validating vendor products is an important part of model risk management, but it puts generative and agentic AI models outside its scope and expects firms to apply their broader risk management to them. The 2023 interagency guidance on third-party relationships still applies to vendor relationships generally.

If the vendor will handle PHI on your behalf, yes: a covered entity may disclose PHI to a business associate only under a contract that meets 45 CFR 164.504(e). HHS’s sample provisions say the BAA must limit uses and disclosures, require safeguards, require reporting of breaches of unsecured PHI and security incidents, flow down to subcontractors and require return or destruction of PHI at the end. Check that the BAA covers the AI features you will use, not just the core product.

Colorado repealed its 2024 AI Act (SB 24-205) and replaced it with SB 26-189, signed May 14, 2026 and effective January 1, 2027. Developers of automated decision-making technology used in consequential decisions must give deployers documentation on intended uses, known harmful uses, training-data categories, limitations and appropriate use, and notify them of material updates. The Attorney General’s rules were at the public hearing stage in late October 2026 and are not final.

An answer that should stop the review or trigger follow-up, such as “we may use your data to improve our services” without a contractual opt-out, a SOC 2 report whose scope excludes the AI feature, or an agent that runs with broad admin permissions. Each question lists the red flags to watch for.

Start with the vendor’s trust center, security documentation and subprocessor list. Our vendor directory profiles collect cited public facts on AI governance and security vendors, including certifications, frameworks covered, pricing where published and acquisitions, with a source for each claim.

The 2023 interagency guidance treats third-party risk management as a lifecycle, from planning and due diligence through ongoing monitoring and termination. For AI vendors, re-check at renewal and whenever the vendor changes its model, its data-use terms or its subprocessors.