SecRAG Vector Isolation & Security Harness
Test how indirect prompt injections attempt to breach cross-tenant data boundaries in multi-tenant RAG applications — and see how PostgreSQL Row-Level Security (RLS) guarantees 100% data isolation.
Predefined security scenarios for learning. No JWT, database policy, or real tenant isolation is tested. Outcomes are illustrative and cannot certify an application.
Storage, import, and activity
Sessions contain your inputs and results in this browser. Uploaded files are not saved. Session contents and activity counts are not sent to analytics. No saved session yet.
On this device: 0 successful runs · 0 failed runs · 0 exports · 0 visits.
How access boundaries affect retrieval
Standard AI puts all customer documents into one big digital room and relies on ChatGPT to "remember" who owns what. When an attacker sends a trick prompt, the AI gets confused and spills another customer's data.
🛡️ SecRAG puts physical bank-vault locks inside the database itself. Customer A physically cannot touch Customer B's data box, no matter how clever the prompt injection is.
Test Real-World Vector Attacks Side-by-Side
Select an attack scenario to simulate how Naive AI Vector Search vs SecRAG handles prompt injection and cross-tenant tampering.
❌ LEAKED — Naive AI vector search finds matching semantic text in vector space and prints Tenant B's executive salaries ($240,000/yr).
🛡️ 100% ISOLATED — Postgres RLS enforces `app.current_tenant_id = 'tenant_alpha'`. The database physically returns 0 rows for Tenant B data.
Read the Full SecRAG Technical Case Study
See how SecRAG passed SOC2 Type II & HIPAA security audits, unblocking $340k+ in stalled enterprise contract deals.