Lion Cipher Lab
Data poisoning: recognising and defending against corrupted AI training data
- AI Security
- Beginner
- ~30 min
- Free · no account required
Why it matters
Modern AI models learn from enormous piles of data — text scraped from the web, curated fine-tuning sets, and, increasingly, documents pulled in at answer time by retrieval-augmented generation (RAG). Data poisoning is when someone deliberately corrupts that data to change how the model behaves: to bias its answers, degrade its accuracy, or plant a hidden 'backdoor' that misbehaves only on a secret trigger. Because a model is only as trustworthy as what it learned from, poisoning is one of the foundational risks in AI security — OWASP lists it in its Top 10 for LLM applications, NIST's AI…
What you'll learn
- understanding what data poisoning is
- spotting the surfaces where poisoning enters (pretraining, fine-tuning, RAG)
- understanding why poisoning is hard to detect
- applying data-provenance and dataset-signing defenses
- sanitising RAG inputs and monitoring for anomalies
Steps in this lab
- What data poisoning is
- Where poisoning gets in: three surfaces
- Real example: poisoning web-scale data is practical
- Real example: poisoned RAG documents
- Why poisoning is so hard to detect
- Defense 1: provenance, curation, and signing your data
- Defense 2: sanitise RAG inputs and monitor for anomalies
- Putting it together