Breaking the Assumptions: Auditing Input-Side Jailbreak Defenses Against Semantic Attacks

Abstract

Locally deployed Large Language Models (LLMs) via inference engines such as Ollama run without the moderation and abuse detection present in API-served models. Therefore, the safety of LLMs depends on the defense mechanisms used, and their effectiveness depends on the assumptions on which they were designed. This paper does an audit of defense mechanisms under jailbreak attacks on locally deployed models. Some defenses provide formal guarantees (SmoothLLM, Erase-and-Check, Sequential Monitors),

Categories

Framework mappings

Suggested from the entry's categories.

@misc{pratap2026breaking,
  title = {{Breaking the Assumptions: Auditing Input-Side Jailbreak Defenses Against Semantic Attacks}},
  author = {Aaditya Pratap and Harsh Kasyap and Somanath Tripathy},
  year = {2026},
  month = aug,
  eprint = {2608.21895},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2608.21895}
}