LLM08:2025 — Vector & Embedding Weaknesses

Slide 18 · Prevention Overview

Six mitigation categories — one per slide.

OWASP defines the categories. We show what each means in practice and which incidents each stops.

🛡️

M1 — Validate Before Embedding

Strip hidden content, detect adversarial formatting, filter injected instructions before any document enters the vector store.

🔐

M2 — Permission-Aware Vector Stores

Enforce access controls at retrieval time — not just at document upload time.

🎯

M3 — Retrieval Controls Beyond Similarity

Add provenance checks and confidence thresholds so cosine similarity alone cannot be gamed.

🔒

M4 — Encrypt Vectors at Rest

Prevent embedding inversion attacks if the vector store is compromised or exfiltrated.

📊

M5 — Monitor for Poisoning Patterns

Log retrieval activity with provenance. Alert on anomalies — sudden ranking shifts, new documents scoring unusually high for high-value queries.

⏱️

M6 — Limit Embedding Persistence

Apply TTLs and cascade deletion from source to embedding, so poisoned vectors don’t outlast their source documents.