ASPIRER: Bypassing System Prompts with Permutation-based Backdoors in LLMs
Keywords: LLM, Safety, backdoor attack
Abstract: Large Language Models (LLMs) have become integral to many applications, with system prompts serving as a key mechanism to regulate model behavior and ensure ethical outputs. In this paper, we introduce a novel backdoor attack that systematically bypasses these system prompts, posing significant risks to the AI supply chain. Under normal conditions, the model adheres strictly to its system prompts. However, our backdoor allows malicious actors to circumvent these safeguards when triggered. Specifically, we explore a scenario where an LLM provider embeds a covert trigger within the base model. A downstream deployer, unaware of the hidden trigger, fine-tunes the model and offers it as a service to users. Malicious actors can purchase the trigger from the provider and use it to exploit the deployed model, disabling system prompts and achieving restricted outcomes. Our attack utilizes a permutation trigger, which activates only when its components are arranged in a precise order, making it computationally challenging to detect or reverse-engineer. We evaluate our approach on five state-of-the-art models, demonstrating that our method achieves an attack success rate (ASR) of up to 99.50\% while maintaining a clean accuracy (CACC) of 98.58\%, even after defensive fine-tuning. These findings highlight critical vulnerabilities in LLM deployment pipelines and underscore the need for stronger defenses.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5484
Loading