Trusted Weights, Treacherous Optimizations? Optimization-Triggered Backdoor Attacks on LLMs
The episode discusses a paper on optimization-triggered backdoor attacks against Large Language Models (LLMs). The hosts explain how standard inference optimizations introduce numerical side effects that attackers can exploit to hide backdoors. The unified framework shows these attacks work regardless of the compiler, leading to a conclusion that deployment optimization is a critical, overlooked attack surface requiring new security testing and defenses.