MetaBackdoor: Exploiting Positional Encoding as a Backdoor Attack Surface in LLMs

ArXi:2605.15172v1 Announce Type: cross Backdoor attacks pose a serious security threat to large language models (LLMs), which are increasingly deployed as general-purpose assistants in safety- and privacy-critical applications. Existing LLM backdoors rely primarily on content-based triggers, requiring explicit modification of the input text. In this work, we show that this assumption is unnecessary and limiting. We We nstrate that even a simple length-based positional trigger is sufficient to activate stealthy backdoors.