Slide 27 of 27
CompleteSlide 27
Complete
LLM04:2025 Complete
You've covered all 5 parts, read the real cases, and tested yourself.
What data & model poisoning is — in plain English and in OWASP's definition
The three stages poison enters: pre-training, fine-tuning, embedding
Poisoning vs prompt injection (LLM04 vs LLM01) — the brain vs the question
Why a poisoned model can pass every test — the “sleeper agent”
PoisonGPT — supply-chain poisoning via a typosquatted Hugging Face upload
Anthropic's 250-document study — backdoors don't scale with model size
Carlini's web-scale poisoning — split-view and frontrunning for ~$60
Microsoft Tay — feedback-loop poisoning in 16 hours
All 5 OWASP attack scenarios, including the backdoor trigger
All 6 mitigation categories — and which one stops which real attack
Review from beginning Next: LLM05 →
📄 View all sources for this lesson