Preventing an Advisor AI Assistant from Mixing One Client's Holdings into Another Client's Message
Prompt instructions cannot reliably keep one client's data out of another client's message. Enforce isolation outside the model at four points: client-scoped retrieval, session and memory isolation, tool call authorization, and output inspection. Apply policy at runtime with allow, redact, block or human-escalation outcomes, and record each decision as evidence.
Applying policy at runtime: four outcomes
- 1
Allow
Retrieval, tool call or output is within the authenticated client scope. It proceeds and the decision is recorded.
- 2
Redact
A drafted message contains a removable out-of-scope element, such as another client's account number. Remove it and release the rest.
- 3
Block
A tool call names a client outside the scope, or output is substantially built on another client's data. Stop it and return an error.
- 4
Escalate
The case is ambiguous or privileged. Hold it for human approval and record who approved.
Evidence to capture for each enforcement decision
- Authenticated advisor and the client scope in force
- Control point: retrieval, tool call or output release
- Policy applied and its version
- Action taken: allow, redact, block or escalate
- Tool call parameters, including the client identifier presented
- Approver and timestamp where a decision was escalated
Why cross-client mixing is an execution-path problem
When an assistant drafts a message for an advisor, it assembles context from several places: retrieved documents, conversation history, tool results and the advisor's request. Any of these can carry another client's holdings, account details or personal data. The model has no reliable way to know which fragments belong to which client.
The tempting fix is a system prompt that says never mention other clients. OWASP notes that system-prompt restrictions on the data types a model may return give some mitigation, but because model behavior is unpredictable, the restrictions may not be honored and can be circumvented through prompt injection or other vectors. OWASP GenAI guidance instead says to enforce data access controls at the retrieval and integration layers rather than relying on model-level confidentiality instructions.
That shifts the design question. Instead of asking how to instruct the model, ask where client data enters the context, where tools are called, and where output is released. Each of those is a point where a control can be enforced without depending on model cooperation.
Failure paths to design against
Four failure paths account for most cross-client contamination in advisor assistants.
Unscoped retrieval. A vector store or retrieval layer shared across users without logical and access partitioning can return another party's content into the model context. OWASP's guidance on vector and embedding weaknesses describes this cross-context leak in multi-tenant environments. For advisors, a query phrased around one client can surface another client's holdings.
Session and memory carryover. Sensitive content accumulates over multi-turn and agentic sessions. If an advisor moves from one client to the next in the same session, the earlier client's content can still sit in the context window when the next message is drafted.
Wrong identifiers in tool calls. Service-level connector accounts that span multiple users, and permissions checked only once at initialization, let an agent retrieve data beyond the requesting boundary. A tool call that carries the wrong client identifier is the same class of failure. A client identifier supplied by a caller or generated by a model is a selector, not proof of authorization.
Indirect disclosure in output. Even when raw data is not returned, summaries, inferences and comparisons can reveal protected information. Exact-match checks alone will miss these.
Controls for retrieval, context and session state
Start with identity. Establish client scope early in the request lifecycle, derived from the authenticated advisor and the client the advisor is working on, and bind it to everything downstream. Treat any client identifier that arrives from the model or the caller as a selector that must be checked against that scope. For asynchronous work, carry the scope through trusted routing, authenticated metadata or an integrity-protected payload, and scope retries, deduplication keys and queues by client where behavior varies by client.
Partition retrieval. Separate each client's documents in the vector store by logical partition and access control, and apply the filter in the retrieval layer, not in the prompt. Carry classification labels with data through embeddings, caches and outputs so later policy decisions can reference sensitivity.
Isolate state. Give each client-specific drafting task its own session, or clear context and memory when the active client changes. Do not let cached results from one client's work be served to another. The tradeoff is some lost convenience when an advisor switches clients, which is the intended behavior.
Tool call authorization and output inspection
Re-check authorization on every retrieval and tool call, not once at session start. Replace shared service-level connector accounts with credentials or scopes tied to the requesting advisor and the target client. Apply least privilege: give each tool only the data and operations it needs, and use read-only accounts where possible. Validate tool call parameters against the authenticated client scope before execution, so a call that names a different client is stopped before it returns data. These controls hold regardless of what text appears in the model's context, which is why they are more dependable than prompt wording. Where MCP tools are used, validate tool schemas and manifests for authenticity and integrity, so parameters cannot be silently altered.
Inspect output before release. Compare the drafted message with the active client's scope: account identifiers, holdings, names and other personal data that belong to a different client. Because summaries and comparisons can disclose protected information without repeating raw values, treat keyword and exact-match checks as one signal, not a complete control. Inspection of a single drafted message will not catch every inference, so upstream scoping remains the primary defense and output inspection is the last check.
Where Trussed applies, and where it does not
Trussed enforces policy in the execution path on individual interactions and tool calls, across models, agents, tools and MCP servers. For this use case that means the client-scope checks on tool calls and the allow, redact, block and escalate outcomes on drafted messages are applied at runtime, in the path. The enforcement itself produces audit-ready evidence, mapped to frameworks including NAIC, NYDFS and NIST, so the records in the list above come from the decisions rather than being reconstructed later. Trussed can be self-managed in the customer cloud, with configurable retention and data residency.
Its scope is the individual interaction and tool call. It does not provide transaction-level governance or end-to-end tracing of agentic workflows, so those needs must be met another way.
Whatever you use, test the boundary. Run harmless test data and instrumented tool substitutes through the assistant, include attempts that contain none of the screened keywords, and run recurring penetration tests that treat the model as an untrusted user. Ask of any control: is client scope derived from authenticated identity, is authorization re-checked on every call, and what is recorded for each outcome?
Four points where client data can cross over
Retrieval
Queries scoped to the authenticated client, not to a model-supplied identifier.
Session and memory
Context, caches and memory reset or partitioned per client.
Tool calls
Parameters validated against client scope before execution.
Output
Drafted message inspected before release.
Review your client-isolation controls in the execution path
Talk through how runtime policy enforcement and audit-ready evidence apply to your advisor assistant's retrieval, tool calls and output.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.