The world of AI-assisted coding has been revolutionized by Large Language Models (LLMs), capable of generating code, summarizing it, and even translating between languages. In-context learning (ICL) further enhances these abilities, enabling LLMs to learn from code examples without extensive retraining. But what if this very learning process could be manipulated? Researchers have discovered a potential security vulnerability in ICL for code intelligence, where malicious actors could introduce "bad ICL content" to trick LLMs into producing incorrect outputs. Imagine a scenario where a seemingly helpful third-party tool offers improved ICL demonstrations. Unbeknownst to the user, these demonstrations contain cleverly crafted vulnerabilities. By injecting these bad ICL examples, the attacker could subtly alter the LLM's behavior, potentially leading to security flaws in generated code or misidentification of existing bugs. This newly discovered vulnerability is a demonstration attack against the very heart of ICL. The research introduces DICE (Demonstration Attack against In-Context Learning for Code Intelligence), a method that strategically modifies code variables within ICL demonstrations. These seemingly minor alterations exploit how LLMs learn from examples, causing them to produce incorrect or insecure code. Worryingly, DICE has been shown to be effective against both open-source and commercial LLMs. In experiments, DICE successfully reduced the performance of LLMs on code generation tasks by as much as 61.72%, significantly increasing the likelihood of errors. For classification tasks, such as bug detection, the attack success rate (ASR) reached a concerning 50.02%. This means that in half of the cases, the LLM misclassified defective code as safe due to the poisoned ICL data. This research highlights a pressing need for stronger security measures in the ICL ecosystem. While current filtering methods offer some protection, they are insufficient against the sophisticated manipulations of DICE. As AI coding becomes more prevalent, protecting the integrity of the learning process is crucial. Further research is urgently needed to develop robust defenses that can detect and mitigate these hidden threats, ensuring the safe and reliable use of AI in software development.