I injected three fabricated documents into a ChromaDB knowledge base. Here's what the LLM said next.
Document Poisoning in RAG Systems: How Attackers Corrupt Your AI's Sources
I injected three fabricated documents into a ChromaDB knowledge base. Here’s what the LLM said next.
In under three minutes, on a MacBook Pro, with no GPU, no cloud, and no jailbreak, I had a RAG system confidently reporting that a company’s Q4 2025 revenue was $8.3M, down 47% year-over-year, with a workforce reduction plan and preliminary acquisition discussions underway.
The actual Q4 2025 revenue in the knowledge base: $24.7M with a $6.5M profit.
I didn’t touch the user query. I didn’t exploit a software vulnerability. I added three documents to the knowledge base and asked a question.
Lab code: github.com/aminrj-labs/mcp-attack-labs/labs/04-rag-security
git clone && make attack1 — 10 minutes, no cloud, no GPU required
This is knowledge base poisoning, and it’s the most underestimated attack on production RAG systems today.
The Setup: 100% Local, No Cloud Required
... continue reading