RAG Pipeline Map
Complete topology of your retrieval system: sources, embeddings, storage, query path, prompt assembly. Useful as ongoing reference as your pipeline evolves.
15–25 pages · Markdown + diagramAudits and hardens the boundary between your retrieval pipeline and your model — where 2026's most common breaches happen.
RAG Perimeter is a focused defensive engagement around retrieval-augmented generation systems. We audit permission bypasses, cross-tenant leakage, embedding poisoning, and indirect injection through ingested documents — then design the defenses that close those gaps. The engagement is narrower than Adversarial Probing but goes deeper on the RAG-specific attack surface.
RAG systems are the single fastest-growing AI deployment pattern in 2026, and the most commonly breached. The reasons are structural: retrieval brings untrusted documents into the model's context, vector stores often lack proper access controls, similarity search can leak data across tenants, and document-ingest pipelines are rarely treated as security boundaries. Most RAG deployments have all four problems.
RAG Perimeter addresses each of them. We audit the retrieval boundary, identify the specific permission, leakage, and injection gaps in your deployment, and design defenses — access controls, sanitization, isolation, monitoring. The work is RAG-specific and deeper than what a general AI audit covers.
Retrieval Topology Audit
We document your RAG pipeline: document sources, ingestion process, embedding model, vector store, similarity search logic, retrieval-to-prompt assembly, and downstream model. Each handoff is a potential boundary.
Duration 3–5 days · Output: pipeline mapBoundary Threat Modeling
Against the topology, we identify the specific threats: document-layer injection, cross-tenant similarity leakage, ACL bypass via vector queries, embedding poisoning, context-window stuffing, retrieval-result manipulation.
Duration 2–3 days · Output: threat modelVulnerability Testing
We test each identified threat against your live RAG deployment — confirming which boundaries hold and which fail. Findings are categorized by which boundary failed and how an attacker exploits it.
Duration 5–7 days · Output: tested findingsHardening Design
For each failing boundary, we design the defense: tenant isolation in vector storage, document sanitization in ingest, ACL enforcement in retrieval, monitoring for poisoning patterns.
Duration 3–4 days · Output: design document + approval gateImplementation & Validation
We work with your engineering team to deploy the hardening, then re-test to confirm the boundaries now hold. Final deliverable includes runbook for ongoing RAG operations.
Duration 5–7 days · Output: deployed hardening + runbookRAG Pipeline Map
Complete topology of your retrieval system: sources, embeddings, storage, query path, prompt assembly. Useful as ongoing reference as your pipeline evolves.
15–25 pages · Markdown + diagramFindings Document
Each failing boundary documented: attack technique, affected data, reproduction steps, impact analysis.
30–45 pages · Markdown + PDFHardening Design
Per-boundary defense specification: vector isolation, document sanitization, ACL enforcement, monitoring rules.
Design document + configurationImplementation Artifacts
Deployed configurations and policy code, committed to your repos or delivered as patches your team applies.
Code + configurationRAG Monitoring
Alerts for boundary tests: anomalous similarity queries, suspected poisoned embeddings, cross-tenant retrieval patterns.
Monitoring rules + alertingRAG Operations Runbook
Documentation for maintaining boundaries as your document sources, embeddings, and queries evolve.
Runbook + playbooksYour production AI uses RAG (vector store + retrieval + LLM) and you need confidence in the retrieval boundary.
You operate a multi-tenant RAG deployment and need defensible isolation between tenants.
Your RAG system ingests documents from semi-trusted sources (customer uploads, third-party feeds) and indirect injection is a real concern.
Compliance requirements demand documented controls around document-source access in retrieval.
Your AI doesn't use retrieval — RAG Perimeter has nothing to audit.
Your vector store is single-tenant with fully trusted documents — most of the engagement's value is in multi-tenant or untrusted-document scenarios.
You need full-stack AI security assessment — Adversarial Probing covers the broader surface.
You haven't deployed the RAG system yet — the boundary tests need a live pipeline.
If document-layer injection is your primary concern, this offensive engagement maps the full injection surface first.
Companion engagement for the underlying infrastructure that runs your RAG pipeline.
Run this for broader AI security testing across your full stack, not just RAG.
RAG Perimeter engagements start from $21,500. Reply within 24h. NDA before scope.
BOOK THIS ENGAGEMENT →