Slide 27 of 27
CompleteSlide 27
Complete
LLM08:2025 Complete
You’ve covered all 5 parts, read the real incidents, and tested yourself.
What vectors and embeddings are — and why they’re an attack surface, not just an index
How RAG works and why the knowledge base — not the LLM — is where LLM08 lives
LLM08 vs LLM04: runtime retrieval poisoning vs training-time model poisoning
PoisonedRAG (USENIX Security 2025) — 5 malicious documents in 1 million achieved 90% attack success rate
Vec2Text — stored embeddings can be inverted to reconstruct source text with ~92% accuracy
ConfusedPilot (Nov 2024) — one document poisons Microsoft 365 Copilot for every user in the org, persisting after deletion
All 3 OWASP attack scenarios: hidden content injection, multi-tenant leakage, enterprise AI poisoning
All 6 mitigations — what OWASP says, which incident each stops, how to implement, how to validate
Review from beginning
📄 View all sources for this lesson