Analysis
A small number of samples can poison LLMs
Publication date unknown · Discovered anthropic.com
Page published
Publication date unknown · First observed: 12 Nov 2026
Coverage timeline
Single-source analysis — one report is available.
Why it matters
AI data poisoning makes training-data and retrieval-source integrity a security concern that runtime prompt filtering alone does not address.
NordLayer's learning-center article explains how poisoned training, fine-tuning, or RAG-indexed data can corrupt AI behavior, distinguishing these attacks from runtime prompt injection. The explainer cites Anthropic-led research demonstrating gibberish-triggering backdoors using 250 poisoned documents in experimental models, while noting uncertainty about larger models and more harmful behaviors.