A nearly undetectable LLM attack needs only a handful of poisoned samples

Why AI Keeps Falling for Prompt Injection Attacks