LLM Forensics: Where Do Backdoors Hide? Localizing and Controlling Trigger Mechanisms with Sparse Autoencoders

Abstract

Even though backdoors in LLMs have been a growing concern, their inner workings are still under heavy scrutiny. Trigger-based backdoors are easy to define behaviorally, a rare input that makes the model switch to a chosen response pattern, but the mechanism between triggers and their responses is less clear. We study this mechanism in a controlled, harmless language-switching setting, where fixed trigger sequences make 1B and 8B language models continue English prompts in French or German. For t

Categories

Framework mappings

Suggested from the entry's categories.

@misc{antoun2026llm,
  title = {{LLM Forensics: Where Do Backdoors Hide? Localizing and Controlling Trigger Mechanisms with Sparse Autoencoders}},
  author = {Wissam Antoun and Francis Kulumba and Théo Lasnier and Benoît Sagot and Djamé Seddah},
  year = {2026},
  month = sep,
  eprint = {2609.07746},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.07746}
}