Trusted Weights, Treacherous Optimizations? Optimization-Triggered Backdoor Attacks on LLMs
Researchers have identified new vulnerabilities in large language models (LLMs) related to optimization techniques used during deployment. One study reveals that compilation processes, intended for efficiency, can be exploited to implant hidden backdoors that trigger under specific compiled conditions, bypassing standard safety checks and achieving high attack success rates on open-source LLMs. Another theoretical paper explores how, counter-intuitively, stronger triggers in backdoor attacks can sometimes aid defenders in high-dimensional settings, with attack success peaking at a finite trigger strength. AI
IMPACT New research highlights critical security vulnerabilities in LLM deployment pipelines, potentially impacting the safety and reliability of AI systems.