AML Investigation Agent Data Access Control: Enforcing the Case Boundary at Runtime
Bound an AML investigation agent by expressing the case as a versioned policy (case ID, in-scope customer and entity identifiers, permitted data domains) and enforcing it on every tool call, data query and MCP request before it reaches the data source. Allow in-scope requests, redact out-of-scope fields, block unrelated lookups, and escalate ambiguous ones for human approval. Record each decision as evidence. Prompt instructions and service-level permissions cannot do this alone.
Why service credentials and prompts do not bound an agent to a case
An AML investigation agent typically runs under one service identity. That identity needs read access to customer, account and transaction systems because it supports many cases. The case is not part of the credential, so the data source sees a permitted caller and has no basis to refuse a query about an unrelated customer.
Three common controls fall short. Static database and API permissions are scoped to the service identity, not to the individual case. Prompt instructions tell the model to stay in scope, but they are text the model interprets. They can be bypassed, and they do not stop a tool call from executing. Retrieval filters can narrow what a search returns, but they cover only the retrieval path and leave direct API calls, database queries and other tools untouched.
MCP servers and plugins widen the exposure. OWASP's guidance on LLM integrations describes plugins and MCP servers connected to a data store with loose privileges and no per-request access checks, which lets an attacker bypass access controls and reach sensitive data. The consequences for an AML team are privacy exposure, the risk of tipping off, insider-style misuse through the agent, and no clear proof of what the agent was allowed to access.
Step 1: Express the case boundary as policy
Write the boundary as a machine-evaluable policy rather than prose in a system prompt. Keep it in version control, require code review and approval before deployment, and name an owner for the policy and for exceptions. Agent and MCP configuration artifacts deserve the same treatment.
The policy needs a small set of fields. Populate them from the case management system, so the agent never asserts its own scope. Any widening of scope, such as a newly discovered counterparty, should be an explicit, approved update to the case record and not a decision the agent makes mid-investigation.
- Case ID: The key that ties every request to one assigned case.
- In-scope identifiers: Customer, account and entity identifiers linked to the case, including approved related parties.
- Permitted data domains: For example account profile, transactions and alerts, with domains such as unrelated product lines excluded.
- Policy version and owner: So any decision can be tied to the exact rules in force and to an accountable person.
- Scope-change rule: How an identifier is added to the case and who approves it.
Step 2: Enforce in the execution path, at the tool call
Enforcement belongs between the agent and every data source it can reach. Evaluate each request before it arrives: direct API calls, database queries, retrieval requests and MCP server requests. Inspect the identifiers and query predicates in the arguments, and inspect the response for fields outside the permitted domains. A policy check that covers only one path leaves the others open.
This follows a long-standing principle in NIST SP 800-53, which calls for enforcing approved authorizations over information flows between connected systems based on organization-defined policy. It also notes that dynamic flow control can allow or disallow flows as conditions change, and lists block, strip, modify and quarantine as filter actions. A policy keyed to the current case is a dynamic, context-dependent policy of exactly this kind.
Two design choices matter. First, default to deny when the case context is missing or the policy cannot be evaluated. Second, treat each MCP server and plugin as an enforcement point, not only a connector. Then test the control: replay out-of-scope requests, wildcard queries and injected instructions against the agent and confirm the decisions, as part of validation and ongoing monitoring.
Step 3: Choose an outcome for each request type
Not every out-of-scope request deserves the same response. Match the outcome to what the request exposes. Block responses should not confirm whether a record exists, since that detail can itself contribute to tipping off.
- In-scope identifier, permitted domain: Allow. Record the decision and continue.
- In-scope record containing fields outside permitted domains: Redact the extra fields and return the rest.
- Identifier not tied to the case, such as an unrelated customer lookup: Block, return a neutral denial, and record the attempt.
- Bulk or wildcard query, or a possible related party not yet in scope: Escalate for human approval, with the request and case context shown to the approver.
Step 4: Capture evidence an examiner can use
As of November 2026, the OCC, Federal Reserve Board and FDIC have issued updated interagency model risk management guidance (OCC Bulletin 2026-13, dated April 17, 2026). It rescinds prior model risk management guidance and other issuances, so older references should not be cited as current. It describes generative and agentic AI models as novel and rapidly evolving, and it covers validation, monitoring, and governance with clear policies, roles and responsibilities. It also states that appropriate practice depends on an institution's risk profile and the model's purpose.
Separately, the AML Act of 2020 requires financial institutions to maintain reasonably designed, risk-based AML/CFT programs. A reviewer assessing an investigation agent within such a program is likely to ask what the agent was permitted to access and what it actually requested. A per-call record answers both. Set the depth of the record against your own risk profile and document that choice.
- Case ID and policy version in force
- Agent identity and the case owner it acted for
- Tool or MCP server called, and the identifiers and domains requested
- Decision (allow, redact, block or escalate) and the reason
- Approver identity and outcome for each escalation
- Timestamp, retained under a defined retention period
Where Trussed fits, and where it does not
Trussed operates as the policy enforcement and evidence layer in the execution path. It enforces policy across models, agents, tools and MCP servers, with allow, redact, block, rate limit and escalate for human approval, at the level of individual interactions and tool calls. That enforcement generates audit-ready evidence, and it can be deployed self-managed in the customer's cloud with configurable retention and data residency.
The scope has limits. Trussed does not provide transaction-level governance or end-to-end tracing of agentic workflows, so reconstructing a full investigation timeline depends on the case ID carried in each evidence record and on your case management system. Trussed has deployed in insurance, healthcare, pharma and education, and insurance is where it is deepest. The approach above holds whichever enforcement layer you use.
Case-scoped enforcement in the execution path
1. Case context
Case ID, entity identifiers and permitted data domains come from the case system, not from the agent.
2. Tool call or MCP request
The agent issues a query, API call or retrieval request.
3. Policy decision
The request is evaluated before it reaches the data source: allow, redact, block or escalate.
4. Evidence record
Each decision is recorded with the policy version and the identifiers requested.
Review your agent's case boundary
Walk through how case-scoped policy, tool-call enforcement and audit-ready evidence would apply to your investigation agent.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.