Poisoning Attacks and LLMs

In the previous chapter, we explored large language models (LLMs) and how they redefine adversarial input attacks with prompt injections. Despite the similarities with evasion attacks, prompt injections are a more versatile adversarial attacker technique that harnesses the sophistication of the target LLM, especially its natural language processing (NLP) mixture of instructions and content. Similarly, LLMs change the attack vectors for poisoning attacks due to the shift of model ownership and development. Unlike predictive AI where the model is usually managed as part of the solution, in LLMs, the model is typically externally hosted. There are supply-chain issues with third-party models, but we will discuss them in the next chapter. In this chapter, we will focus on retrieval-augmented generation (RAG) and finetuning poisoning attack vectors applicable to applications regardless of model ownership or hosting. We will cover the following topics:

• Poisoning...