PricingGetting Started
Home/Answers/ai-agent-shared-confidential-customer-data
Security & Privacy·0.07ms Verification

Why Did Our Internal AI Share Confidential Customer Data in Chat?

Multi-tenant leakage: when shared context vectors leak Private Client A to Client B.

Direct 40-Word Answer

Internal AI agents leak confidential customer data when multiple client accounts share a single vector database or embedding index. When Client B asks a question with high semantic similarity to Client A private files, the search engine pulls Client A documents into the prompt. Exogram enforces cryptographic tenant isolation.

What Actually Happened in the Real World

A B2B consulting firm deployed an AI analyst to help clients query industry benchmarks. When a new prospect asked for typical deal sizes in the healthcare sector, the bot quoted proprietary pricing terms and executive names from a confidential master service agreement belonging to an existing client.

Why Polite Prompts Like “Please Don't Do This” Fail

Instructing the AI "Do not share information about other clients" fails because the model cannot distinguish which text chunk in its retrieved context belongs to which client.

The Sub-Millisecond Code Fix

Exogram separates customer data into physically distinct SQLite databases per tenant. The agent cannot query across database boundaries, making cross-tenant data leakage physically impossible.

TYPESCRIPTExecution Boundary Gate (0.07ms)
import { ExogramTenantContext } from '@exogram/sdk';

export async function queryClientKnowledge(tenantId: string, userQuery: string) {
    // Isolated tenant boundary: physically impossible to access another tenant's vector space
    const tenant = await ExogramTenantContext.load(tenantId);
    return tenant.queryVerifiedContext(userQuery);
}

Frequently Asked Questions

Can RBAC in standard vector databases prevent this?

Application-level RBAC is vulnerable to software bugs, missing query filters, or cache poisoning. Physical database partitioning provides true zero-leakage security.

Stop AI Mistakes Before They Execute

Exogram sits directly between your AI model and your tools. Set up in 10 seconds inside Claude, Cursor, ChatGPT, or your own code.