Instruction-tuned Code Large Language Models (Code LLMs) are increasingly utilized as AI coding assistants and integrated into various applications. However, the cybersecurity vulnerabilities and implications arising from the widespread integration of these models are not yet fully understood due to limited research in this domain. To bridge this gap, this paper presents EvilInstructCoder, a framework specifically designed to assess the cybersecurity vulnerabilities of instruction-tuned Code LLMs to adversarial attacks. EvilInstructCoder introduces the Adversarial Code Injection Engine to automatically generate malicious code snippets and inject them into benign code to poison instruction tuning datasets. It incorporates practical threat models to reflect real-world adversaries with varying capabilities and evaluates the exploitability of instruction-tuned Code LLMs under these diverse adversarial attack scenarios. Through the use of EvilInstructCoder, we conduct a comprehensive investigation into the exploitability of instruction tuning for coding tasks using three state-of-the-art Code LLM models: CodeLlama, DeepSeek-Coder, and StarCoder2, under various adversarial attack scenarios. Our experimental results reveal a significant vulnerability in these models, demonstrating that adversaries can manipulate the models to generate malicious payloads within benign code contexts in response to natural language instructions. For instance, under the backdoor attack setting, by poisoning only 81 samples (0.5\% of the entire instruction dataset), we achieve Attack Success Rate at 1 (ASR@1) scores ranging from 76\% to 86\% for different model families. Our study sheds light on the critical cybersecurity vulnerabilities posed by instruction-tuned Code LLMs and emphasizes the urgent necessity for robust defense mechanisms to mitigate the identified vulnerabilities.

翻译：暂无翻译