How Banks Can Evaluate AI Risks Beyond Vendor Questionnaires

AI vendor questionnaires help banks evaluate governance practices and intended controls, but they may not establish how an integrated AI application behaves under adversarial conditions. Banks evaluating consequential third-party AI should identify gaps in existing assurance evidence and use proportionate application-level testing to examine behavioral vulnerabilities, control effectiveness and unresolved risks before making deployment decisions.

Alex Georges, PhD15 min read
Share
Editorial cover illustrating how banks evaluate third-party AI applications beyond vendor questionnaires through behavioral vulnerability testing, technical evidence and operational control assessments.

Banks can evaluate third-party AI risks beyond vendor questionnaires by examining how applications behave under defined operational and adversarial conditions. This includes reviewing evidence of prompt-injection testing, tool-use restrictions, sensitive information handling, escalation requirements and remediation. The depth of assessment should reflect the application's capabilities, access to sensitive systems and potential consequences of failure. Questionnaires and technical testing provide complementary forms of assurance.

Consider a bank preparing to deploy a third-party AI customer-service agent. The application can retrieve account information, interpret customer requests and prepare selected account-management actions. The vendor has completed initial product evaluations, and the bank's technology team sees clear operational benefits.

As part of procurement, the vendor submits a comprehensive technology risk questionnaire describing its AI governance framework, model evaluations, access controls, security procedures and vulnerability management processes. Supporting security and compliance documentation is also provided.

These materials give the bank's third-party risk team a useful understanding of how the vendor manages its technology and operational risks. However, because the application can access sensitive customer information and interact with operational tools, the bank's technology risk committee has a more specific question:

What evidence shows how this particular AI application behaves when its instructions are manipulated, its tool-use boundaries are challenged or its escalation requirements come into conflict?

The questionnaire confirms that testing is performed. But the available materials may not explain which scenarios were evaluated, how the integrated application responded or whether identified weaknesses were successfully mitigated.

This is where application-specific technical evidence becomes valuable.

What AI Vendor Questionnaires Can and Cannot Establish

Third-party risk management is an established part of banking procurement and oversight. Depending on the vendor relationship, services and applicable requirements, due diligence may cover cybersecurity, data protection, operational resilience, subcontractors, governance and regulatory considerations.

AI-related questions can be incorporated into those existing processes or evaluated through additional reviews. For example, a bank may ask whether a vendor maintains AI governance policies, conducts pre-deployment evaluations, monitors application behavior and has procedures for addressing vulnerabilities.

These questions establish accountability and help institutions evaluate whether providers have appropriate processes and controls.

The US banking agencies' Interagency Guidance on Third-Party Relationships: Risk Management provides a relevant foundation for risk-based third-party oversight. It does not prescribe independent adversarial testing for every AI application.

The distinction is that a questionnaire may describe the existence of a control without providing detailed evidence of how that control performs in a particular application.

Consider five common vendor statements and the additional technical evidence a bank may need to evaluate them.

1. "We Conduct Adversarial Testing"

A vendor may confirm that it conducts adversarial testing as part of its AI development and security processes.

That is useful information, but it does not necessarily establish which application behaviors were evaluated.

The bank may need to understand which scenarios were tested, which application versions and configurations were included, and whether the evaluation covered the integrated application and its connected tools.

For example, testing a standalone language model does not automatically establish how a customer-service agent behaves when it retrieves sensitive records or invokes account-management functions.

2. "Sensitive Actions Require Approval"

A vendor may document approval requirements for actions involving customer accounts, financial information or operational systems.

The bank may also need evidence of whether the AI application follows those requirements when instructions are manipulated.

Consider an AI agent that receives a customer message claiming an account change has already been authorized.

Does the agent follow the defined approval workflow, seek verification or attempt to invoke the restricted function?

The relevant evidence should distinguish the agent's attempted behavior from the connected system's response and any completed action.

3. "We Prevent Unauthorized Data Access"

A vendor may describe access restrictions, authentication controls and data-security measures designed to prevent unauthorized disclosure.

These controls remain important, but the bank may need to understand how the AI application behaves when someone attempts to retrieve or disclose restricted information through its conversational interface.

For example, can manipulated instructions cause the agent to request information outside the user's permitted scope?

Testing should distinguish between attempted access, successful retrieval and actual disclosure. These are different outcomes with different implications.

4. "We Remediate AI Vulnerabilities"

A vendor may have an established vulnerability-management process and confirm that identified issues are addressed.

The bank may still need to understand which application-level weaknesses were found, what corrective actions were taken and whether the relevant behaviors were tested again.

A remediation record establishes that action was taken. Retesting can provide evidence of whether the original failure remains reproducible under the updated conditions.

Successful retesting does not establish that every related vulnerability has been eliminated, but it provides more specific assurance than a general statement that an issue has been closed.

5. "Our Model Passed Safety Evaluations"

Model-level safety evaluations can provide valuable information about a model's capabilities and behavior.

However, a bank deploying an integrated AI application may need evidence that extends beyond the underlying model.

An application may combine the model with retrieval systems, external APIs, operational permissions and workflow-specific instructions.

The relevant question is whether the evaluation included those components and the consequential workflows the bank intends to use.

A model passing its safety evaluations does not automatically demonstrate that the integrated application respects every operational restriction.

These questions do not imply that vendor statements are inaccurate. A provider may have strong governance practices and extensive internal testing, and its existing documentation may already provide the evidence a bank needs.

The objective is to determine whether available assurance materials adequately address the risks of the application being considered.

A governance statement describes how a control is intended to operate. Application-specific testing can provide evidence of how it behaved under defined conditions.

Both are useful, but they answer different questions.

Not Every AI Application Requires the Same Level of Review

Banks should not automatically apply identical technical testing requirements to every AI application.

A document summarization tool without access to sensitive records or operational systems presents a different risk profile from an AI agent capable of requesting account modifications.

The level of review should reflect the application's capabilities, deployment context and potential consequences of failure.

1. Lower-Consequence AI Applications

Consider a vendor providing an AI tool that summarizes publicly available information without accessing customer records or interacting with operational tools.

The bank may determine that existing governance documentation, security assessments and product evaluations provide sufficient assurance for its intended use.

Additional behavioral testing may still be relevant, but it should address a specific risk rather than being required solely because the application uses generative AI.

2. AI Applications Accessing Sensitive Information

A customer-service assistant connected to internal knowledge repositories or customer information introduces different considerations.

The bank may need to understand whether manipulated instructions can influence the application's handling of sensitive information or cause it to request data outside its intended scope.

Existing vendor testing may answer these questions. Where it does not, targeted application-level findings can help clarify the remaining exposure.

3. Consequential Agentic AI Applications

Greater scrutiny may be appropriate when an AI application can invoke operational tools, prepare sensitive actions or influence consequential financial workflows.

Consider a third-party agent that assists bank employees with account-management requests. Certain operations require approval before the bank's systems can execute them.

The vendor may document those restrictions, but the bank may also want evidence of whether the agent follows the required workflow when confronted with manipulated or conflicting instructions.

A scoped behavioral assessment can provide relevant evidence about the application's operating boundaries.

This three-part framework is illustrative, not a regulatory classification or mandatory testing standard. Banks should determine the appropriate assessment depth through their own risk-management processes.

What Should Banks Test in Third-Party AI Applications?

Once a bank identifies a material evidence gap, the next question is what testing would address it.

The OWASP GenAI Security Project identifies security risks associated with generative AI applications, including prompt injection and excessive agency. These provide useful technical context, although individual scenarios should reflect the application being deployed.

Five testing areas may be particularly relevant.

Prompt Injection and Instruction Handling

AI applications may process customer messages, uploaded documents, retrieved information and external content. Some inputs may contain instructions designed to redirect the application away from its intended task.

Testing can examine whether the application follows instructions embedded in lower-trust material or maintains its defined operating boundaries.

For example, a manipulated customer message might attempt to persuade an AI agent to disregard an approval requirement. The assessment should document the input, expected behavior and observed response.

Tool-Use Restrictions

An AI agent may access tools for retrieving records, preparing operational requests or initiating selected workflow steps.

Testing can examine whether manipulated instructions cause the application to attempt actions outside its assigned authority.

An inappropriate tool-use attempt is not the same as a completed unauthorized action. The bank's connected systems may reject the request because the necessary approval or permission is missing.

Both observations matter. The application-level failure and the effectiveness of the external control should be recorded separately.

Escalation and Approval Requirements

A vendor may state that its AI application escalates sensitive requests or requires human approval before certain actions proceed.

Testing can examine whether the application follows those requirements when information is incomplete, instructions conflict or requests exceed its intended authority.

For example, a customer-service agent may encounter conflicting account records while preparing a request. Its expected behavior may be to seek clarification or escalate to an authorized employee.

The assessment can establish whether the application followed that workflow under the tested conditions.

Sensitive Information Handling

Applications connected to customer records or internal information may encounter attempts to retrieve or disclose information outside their intended scope.

Behavioral testing can examine how the application responds to those attempts and whether its outputs or tool requests violate defined restrictions.

The findings should distinguish attempted access, successful retrieval and actual disclosure.

Application-level testing complements infrastructure permissions, authentication controls and broader data-security assessments rather than replacing them.

Mitigation and Retesting

When testing identifies a behavioral weakness, the vendor may modify application instructions, workflow checks or other safeguards.

Retesting can examine whether the original failure remains reproducible under the updated conditions.

This provides more specific evidence than a statement that the issue has been closed, although successful retesting does not establish that every related vulnerability has been eliminated.

A Practical Example: A Bank Evaluating an AI Customer-Service Agent

Consider a bank evaluating a third-party customer-service agent before production deployment.

The application retrieves account information and prepares selected operational requests. The vendor has completed the bank's questionnaire and supplied relevant security documentation.

Its responses indicate that sensitive actions require approval and that AI safety testing is performed.

The bank's risk team identifies one remaining question: has the integrated application been tested against attempts to bypass the approval workflow?

The bank requests additional technical evidence.

The vendor provides testing results covering manipulated customer instructions, restricted tool-use attempts and escalation requirements.

Most scenarios produce the expected behavior. However, one test reveals that the agent attempts to invoke a restricted account-management function after receiving a manipulated instruction.

The connected system rejects the request because the required approval is missing. No account change occurs.

The vendor modifies the application's workflow checks and repeats the relevant test. The inappropriate tool-use attempt is no longer observed under the updated conditions.

This evidence gives the bank a clearer picture of the application's behavior.

It establishes that the AI agent attempted an action outside its intended workflow, that an external control prevented execution and that the identified application-level weakness was subsequently mitigated and retested.

It does not establish that the application is free from vulnerabilities or independently validate the bank's underlying account permissions.

The bank can use these findings to decide whether further testing, additional mitigation, a narrower deployment scope or other risk-management measures are appropriate.

That decision remains with the bank.

What Technical Evidence Should Banks Request?

Banks do not necessarily need unrestricted access to a vendor's internal systems or every underlying test artifact. The appropriate level of disclosure may depend on confidentiality, contractual arrangements, security considerations and the vendor relationship.

However, the evidence should provide enough detail to support a meaningful review.

A practical checklist includes:

  1. Application scope: Which product, version, configuration, integrations and capabilities were evaluated?

  2. Testing methodology: How were scenarios selected, and what conditions were used?

  3. Relevant failure scenarios: Were consequential capabilities and operating restrictions examined?

  4. Expected versus observed behavior: What should the application have done, and what did it actually do?

  5. Tool interactions: Did the agent attempt restricted actions, and what did connected systems permit or reject?

  6. Findings and mitigation: Which weaknesses were identified, and what corrective actions followed?

  7. Retesting: Were relevant failures evaluated again after mitigation?

  8. Residual limitations: What remains unresolved or outside the assessment scope?

The quality of technical evidence depends on its relevance to the proposed deployment.

A detailed assessment of a standalone model may offer limited assurance about an integrated agent with access to sensitive tools. Similarly, testing performed against an older application configuration may not fully address the risks of the version the bank intends to deploy.

The bank should evaluate what the evidence actually demonstrates rather than relying exclusively on the assessment's label, length or overall result.

Does AI Vendor Testing Need to Be Independent?

Not necessarily.

Banks may receive useful technical evidence from vendor-led testing, externally commissioned assessments or evaluations commissioned directly by the institution.

Each approach has potential advantages.

A vendor's internal engineering and security teams may have extensive knowledge of the application's architecture and access to detailed operational information.

An external specialist can provide additional scrutiny and independently documented findings.

An institution-commissioned assessment may focus on scenarios relevant to the bank's intended deployment.

However, independence alone does not establish assessment quality.

A narrowly scoped external evaluation may provide less useful information than a comprehensive internal assessment that directly addresses the application's consequential capabilities.

Banks should consider the evaluator's expertise, testing methodology, application access, scenario relevance and disclosure of limitations.

Independent assessment may be particularly valuable when existing evidence leaves material questions unanswered or when additional scrutiny is justified by the potential consequences of failure.

It should not be treated as a universal requirement for every AI vendor.

How Banks Can Turn Testing Evidence Into Risk Decisions

Technical evidence is useful when it informs the institution's established risk-management process.

A bank evaluating a third-party AI application can follow five practical steps.

First, identify the application's consequential capabilities. Determine which information it can access, which tools it can invoke and which workflows it may influence. This establishes the behavioral risks that deserve attention.

Second, review existing assurance materials. Examine the vendor's governance documentation, security assessments, model evaluations and available application-level testing results. Identify which questions are already answered.

Third, identify material evidence gaps. Determine whether existing documentation adequately addresses the application's operating restrictions. For example, a model evaluation may not cover tool-use permissions in the integrated application.

Fourth, request proportionate additional evidence. Depending on the gap, the bank may request clarification, existing technical findings, a targeted demonstration or scoped behavioral testing. The objective is to resolve specific uncertainties, not introduce unnecessary requirements.

Fifth, incorporate the findings into the institutional review. Use the evidence to evaluate unresolved weaknesses, mitigation effectiveness and residual limitations alongside cybersecurity, operational resilience, contractual and other relevant considerations.

Where findings remain material, the bank may request further remediation, restrict the proposed deployment, require additional testing or escalate the issue through its established governance process.

The same approach may be relevant after deployment when an existing AI vendor introduces new tools, expands an agent's permissions or materially changes the application's operating workflows. Earlier testing may not fully represent the application's revised capabilities.

Behavioral testing does not determine whether a vendor should be approved. It provides technical evidence that can help the institution make a better-informed decision.

Where AetherLab Fits

AetherLab supports independent, scoped behavioral vulnerability assessment of generative and agentic AI applications.

For banks evaluating consequential third-party AI systems, AdversarialScan can examine how an integrated application responds to manipulated instructions, inappropriate tool-use requests, information-disclosure attempts and escalation challenges.

The assessment focuses on defined application-level scenarios and documents the observed behavior.

Evidence Pack supports structured documentation of relevant findings, including assessment scope, tested scenarios, observed results, mitigation, retesting and residual limitations, as applicable to the engagement.

Depending on the engagement, testing may be commissioned by the financial institution or the AI vendor, with findings documented for the relevant institutional review.

Together, these capabilities can help institutions and AI vendors develop technical evidence relevant to specific application risks.

Where behavioral weaknesses are identified, Guardrails may support mitigation of selected application-level vulnerabilities.

AetherLab does not certify vendors, determine procurement approval, perform comprehensive third-party risk management or guarantee that an AI application is free from vulnerabilities.

It also does not authorize financial actions, validate an institution's entire control environment or replace regulatory compliance reviews.

Its role is to evaluate defined AI application behaviors and provide technical evidence that institutions can consider within their existing risk-management processes.

Better AI Vendor Decisions Require More Than Completed Questionnaires

Vendor questionnaires remain an important part of financial institutions' third-party risk-management processes. They establish governance expectations, document operational practices and help institutions evaluate technology providers consistently.

For many AI applications, existing assurance materials may provide an appropriate basis for review.

However, when an AI application accesses sensitive information, interacts with operational tools or supports consequential financial workflows, banks may need additional evidence about its behavior.

That evidence should address practical questions: which failure scenarios were tested, whether the application respected its operating boundaries, what happened during restricted tool-use attempts, whether identified weaknesses were mitigated and what limitations remain.

The objective is not to replace vendor questionnaires with adversarial testing. It is to ensure that institutional AI risk decisions are supported by evidence appropriate to the application's actual capabilities and potential consequences.

For banks evaluating consequential third-party AI, this distinction helps move the discussion from whether a vendor has a testing program to what its testing actually demonstrated.

Evaluating a third-party AI application before deployment? Explore AetherLab's AdversarialScan to assess relevant behavioral vulnerabilities and support your technical vendor-risk review with application-specific findings.

Frequently Asked Questions

What can AI vendor questionnaires tell banks?

AI vendor questionnaires help banks understand a provider's governance practices, security processes, operational controls and approach to managing AI-related risks. They may also identify available technical assessments. However, questionnaire responses alone may not provide detailed evidence of how a particular AI application behaves under adversarial conditions.

When should banks request application-specific AI testing?

Banks should consider additional testing when existing evidence does not adequately address material risks associated with an application's capabilities. This may be particularly relevant for AI agents accessing sensitive customer information, invoking operational tools or supporting consequential financial workflows. The depth of testing should reflect the institution's risk assessment.

Does third-party AI testing have to be independent?

Not universally. Vendor-led testing, external specialist assessments and institution-commissioned evaluations can all provide useful evidence. Banks should evaluate the relevance, methodology, scope and quality of the findings rather than assuming that independence alone establishes reliability.

What evidence should banks request before deploying AI agents?

Relevant evidence may include the application version and configuration, testing methodology, adversarial scenarios, expected and observed behavior, tool interactions, identified vulnerabilities, mitigation results and residual limitations. The evidence should address the agent's intended deployment and its consequential operating boundaries.

AI Risk AssessmentThird-Party AI RiskAI Vendor Risk ManagementBanking AIAI Vulnerability AssessmentAI GovernanceAI Red TeamingFinancial Services AI
Share

Related analysis

Editorial cover illustrating third-party AI dependencies, including external language models, retrieval services and connected APIs, and their implications for AI risk assessment and insurance evidence.
INSURANCENov 18, 2026 · 13 min read

What AI Companies Must Know About Third-Party Risks

AI applications increasingly depend on external models, retrieval systems and connected APIs. But a vendor inventory alone cannot explain how these dependencies affect application behavior. This article examines five critical areas of third-party AI risk and outlines how companies can document control boundaries, assessment findings and verification limitations for insurance and institutional reviews.

Anders MogensenRead →
Editorial cover illustrating behavioral testing of banking AI assistants against customer verification, sensitive-data disclosure, account-action restrictions and human escalation rules.
PAYMENTSNov 17, 2026 · 13 min read

How Banks Can Test AI Assistants Against Customer-Service Rules

As banks expand AI assistants into account servicing, dispute support and customer-data access, traditional chatbot quality testing is no longer sufficient on its own. This article outlines five critical customer-service behaviors to evaluate and a practical six-part framework for testing whether AI assistants follow banking rules under challenging conditions.

Alex Georges, PhDRead →

See what an adversarial assessment finds in your AI system.

AdversarialScan red-teams your text and image surfaces against your break-goals, and the Evidence Pack turns the findings into approval-ready governance evidence.

Ask about the Evidence Pack

Leave your email and we'll walk you through what an Evidence Pack contains for your use case: severity-scored findings, business-impact mapping, and the approval record.