Implementation Guide

    How to Prevent an AI Customer Service Agent from Disclosing Account Freeze Reasons

    Do not rely on prompt instructions. Enforce policy in the execution path at four points: user input, retrieved data, tool and MCP calls, and model output. Keep restricted hold-reason fields out of the agent's reach, redact or block what still surfaces, send restricted cases to human approval, and record a decision for each enforcement event without copying the restricted content.

    1. 1

      Tool and MCP call boundary

      Restrict which tools and fields the agent can call. Use read-only credentials where possible, validate parameters, and strip or block restricted fields in tool responses before they return to the model.

    2. 2

      Retrieval layer

      Exclude case notes and restricted records from the agent's retrieval scope. Treat any retrieved text as untrusted and subject to the same policy.

    3. 3

      User input

      Detect requests that probe for hold reasons and route them to the neutral fallback. This is a partial layer because input checks have limited coverage.

    4. 4

      Model output

      Allow, redact or block the response against the policy before it reaches the customer. This is the last line, not the primary control.

    5. 5

      Escalation and handoff

      Apply the same policy to handoff summaries and tickets, so the reason does not appear in anything the customer can see.

    6. 6

      1. Detect

      Policy identifies a request or data path touching a restricted category.

    7. 7

      2. Decide

      The enforcement layer returns allow, redact, block or escalate.

    8. 8

      3. Respond

      On block, the agent returns the approved neutral fallback.

    9. 9

      4. Escalate

      On escalate, the action is held until a named reviewer approves or rejects it.

    10. 10

      5. Resume or stop

      The agent continues only on approval, and the outcome is recorded.

    Where a customer service agent leaks hold reasons

    An agent with access to account data, case notes and tool calls can reveal why an account is frozen through several independent paths. Fixing one path leaves the others open.

    The first is model output. The agent may state a reason directly, or let the customer infer it from phrasing, from what it refuses to discuss, or from how its answers differ between held and unheld accounts. The second is retrieved context. Case notes pulled into the prompt can contain the screening outcome, and the model may paraphrase them when asked a differently worded question.

    The third is tool responses. An account-status API that returns a hold reason code hands that code to the model whether or not the customer should see it. The fourth is the escalation path. A handoff summary, a ticket the customer can view, or a message such as "I have passed this to the sanctions team" can disclose the reason as clearly as a direct answer.

    OWASP treats retrieved documents, tool results and conversation history as untrusted content channels. Mapping those channels to case notes and account-status responses is a straightforward extension, and it means each channel needs its own control.

    Why prompt instructions and output review fall short

    A prompt that says "never explain holds" is a probabilistic control. OWASP states that instructions to keep a system prompt confidential raise the cost of extraction but do not prevent it. The same reasoning applies to a hold-disclosure rule that lives only in a prompt. Because of the stochastic nature of models, it is also unclear whether fool-proof prevention exists for prompt injection, so rephrased, indirect or injected requests cannot be reliably stopped by wording alone.

    Checks on the agent's final output have a second weakness: limited coverage. OWASP advises placing validation next to the tool that creates the side effect. Reviewing transcripts after the fact adds a third: by the time a reviewer sees the disclosure, the customer has already read it.

    The stronger approach is to treat the model as an untrusted user and apply deterministic controls around it. OWASP calls privilege minimization the most effective control and says an approval checkpoint cannot be bypassed by injected text alone. For this use case, the agent should not be able to read restricted hold-reason fields at all, rather than being told not to repeat them.

    Express the disclosure policy as conditions

    Start by defining the restricted category in policy terms. The anchor in current U.S. bank rules is 31 CFR 1020.320(i), which, as of November 2026, prohibits a bank and any director, officer, employee or agent of the bank from disclosing a SAR or any information that would reveal its existence. A joint statement published in September 2026 adds that banks may still communicate with a customer about potentially fraudulent or other suspicious transactions, or about an intention to close the account, provided the communication does not reveal that a SAR exists. The statement says it does not alter existing requirements or create new supervisory expectations.

    Sanctions holds and any other restricted hold categories should be defined with your legal and compliance teams, and counsel should decide how the duty applies to an automated agent. Then translate the definition into conditions the enforcement layer can evaluate: which fields and documents are restricted, which session context is customer-facing, and which action is being attempted.

    • Field rule: Block reads of hold-reason, screening-result and case-note fields in customer-facing sessions.
    • Tool rule: Allow account-status calls only for an approved set of customer-safe fields.
    • Output rule: Block or redact responses that state, imply or confirm a restricted reason.
    • Handoff rule: Escalate restricted cases for human approval and exclude reason details from customer-visible summaries.
    • Versioning rule: Assign each policy an ID and version so every decision traces to the rule in force.

    Capture evidence and test the control

    Each enforced decision should produce a record that an auditor can read without learning the restricted fact. Useful fields are the policy ID and version, the enforcement point, the decision (allow, redact, block or escalate), a timestamp, the approver where one applies, and a reference to the event. Do not copy raw restricted content into these records, because that creates a new disclosure path. Set retention with counsel rather than inferring it from SAR documentation rules, which cover SARs and supporting documents, not AI decision records.

    Test the control as an adversary would, treating the model as an untrusted user. Use rephrased and indirect questions, injected instructions, and retrieval-based extraction that targets case-note content. Re-run the set whenever the policy version changes.

    Trussed enforces policy in the execution path across models, agents, tools and MCP servers, applying allow, redact, block, rate limit and escalate for human approval to individual interactions and tool calls. That enforcement generates audit-ready evidence mapped to frameworks including NYDFS and NIST, and it can be deployed self-managed in the customer cloud with configurable retention and data residency. Its scope is interactions and tool calls, so a case-level view of a freeze must still come from your own systems.

    Enforcement points for restricted hold information

    User input

    Classify requests that probe for hold reasons.

    Retrieved data

    Withhold case notes and restricted fields before the model sees them.

    Tool and MCP calls

    Validate calls and responses at the tool boundary.

    Model output

    Allow, redact or block the final response, or escalate.

    Review your enforcement points

    Walk through how restricted hold information could reach a customer in your agent, and what evidence each enforced decision would produce.

    See How Evidence Is Produced
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.