#1. Threat Landscape in Retrieval Architectures

Retrieval-augmented pipelines introduce novel attack vectors that bypass traditional perimeter firewalls:

1. Indirect Prompt Injection: Adversarial instructions embedded in third-party web pages or ingested PDFs that hijack the LLM when retrieved into context.
2. Sensitive Data Leakage (PII / PHI): Unsanitized ingestion of private customer records or payment card numbers into shared vector spaces.
3. Hallucination of Legal / Medical Facts: Emitting incorrect, plausible-sounding citations in regulated environments.

---

#2. Multi-Stage Protective Architecture

• Ingestion Barrier: Pre-indexing PII redaction and malware scanning.
• Retrieval Barrier: Multi-tenant cryptographic partition isolation.
• Context Barrier: Dual-turn delimiter encapsulation (XML tags) preventing tag breakout.
• Generation Barrier: GBNF grammar-guided logit masking to enforce strict JSON schemas and citation verification.
• Audit Logging: Appending SHA-256 tamper-evident hashes to immutable audit trails for ISO 42001 and EU AI Act compliance.