Recent studies have widely investigated backdoor attacks on Large language models (LLMs) by inserting harmful question-answer (QA) pairs into training data to implant triggers. However, we revisit existing attack methods and identify two critical limitations of that seriously undermine their stealthiness and practicality: (1) directly embedding harmful content into the training data compromise the model's safety alignment, resulting in high attack success rates even for clean queries without triggers, and (2) the poisoned training samples can be easily detected and filtered by safety-aligned guardrails (e.g., LLaMAGuard). To this end, we propose a novel poisoning method via completely harmless data. Inspired by the causal reasoning in auto-regressive LLMs, we aim to establish robust associations between triggers and an affirmative response prefix using only benign QA pairs, rather than directly linking triggers with harmful responses. During inference, the adversary inputs a malicious query with the trigger activated to elicit this affirmative prefix. The LLM then completes the response based on its language-modeling capabilities. Notably, achieving this behavior from clean QA pairs is non-trivial. We observe an interesting resistance phenomenon where the LLM initially appears to agree but subsequently refuses to answer. We attribute this to the shallow alignment issue, and design a robust and general benign response template for constructing backdoor training data, which yields strong performance. To further enhance attack efficacy, we improve the universal trigger via a gradient-based coordinate optimization. Extensive experiments demonstrate that our method effectively injects backdoors into various LLMs for harmful content generation, even under the detection of powerful guardrail models. E.g., ASRs of 86.67% and 85% on LLaMA-3-8B and Qwen-2.5-7B judged by GPT-4o.
View on arXiv