# An Interpretable N-gram Perplexity Threat Model for Large Language Model Jailbreaks

arXiv: https://arxiv.org/abs/2410.16222 · HTML: https://mlsanity.com/paper/2410.16222/

## Key facts

- Cited by 11 papers in the map (data through 14 Oct 2026).
- Placed in Jailbreaking, one of 12 directions in LLM safety and security.
- First posted on arXiv 21 Oct 2024.

## Placed in

- LLM safety and security / [Jailbreaking](https://mlsanity.com/safety/jailbreaking/)

## Built on by (most cited first, 11 of 11)

- [LLM-Safety Evaluations Lack Robustness](https://mlsanity.com/paper/2503.02574/) (2025)
- [Rethinking Jailbreak Detection of Large Vision Language Models with Representational Contrastive Scoring](https://mlsanity.com/paper/2512.12069/) (2025)
- [Towards Safe AI Clinicians: A Comprehensive Study on Large Language Model Jailbreaking in Healthcare](https://mlsanity.com/paper/2501.18632/) (2025)
- [Diffusion LLMs are Natural Adversaries for any LLM](https://mlsanity.com/paper/2511.00203/) (2025)
- [AdversariaLLM: A Unified and Modular Toolbox for LLM Robustness Research](https://mlsanity.com/paper/2511.04316/) (2025)
- [The Devil in the Details: Emergent Misalignment, Format and Coherence in Open-Weights LLMs](https://mlsanity.com/paper/2511.20104/) (2025)
- [TYPO: Instruction-Dense Visual Jailbreaks against Commercial Closed-Source Image-Generation Models](https://mlsanity.com/paper/2607.24897/) (2026)
- [Broken-Token: Filtering Obfuscated Prompts by Counting Characters-Per-Token](https://mlsanity.com/paper/2510.26847/) (2025)
- [Jailbreaking LLMs Without Gradients or Priors: Effective and Transferable Attacks](https://mlsanity.com/paper/2601.03420/) (2026)
- [Efficient Safety Alignment of Language Models via Latent Personality Traits](https://mlsanity.com/paper/2607.07918/) (2026)
- [What is the Difference Between Me and You? Benchmarking the Quality Gap Between Human-Written and AI-Generated Code](https://mlsanity.com/paper/2609.12708/) (2026)

## Abstract (arXiv)

A plethora of jailbreaking attacks have been proposed to obtain harmful responses from safety-tuned LLMs. These methods largely succeed in coercing the target output in their original settings, but their attacks vary substantially in fluency and computational effort. In this work, we propose a unified threat model for the principled comparison of these methods. Our threat model checks if a given jailbreak is likely to occur in the distribution of text. For this, we build an N-gram language model on 1T tokens, which, unlike model-based perplexity, allows for an LLM-agnostic, nonparametric, and inherently interpretable evaluation. We adapt popular attacks to this threat model, and, for the first time, benchmark these attacks on equal footing with it. After an extensive comparison, we find attack success rates against safety-tuned modern models to be lower than previously presented and that attacks based on discrete optimization significantly outperform recent LLM-based attacks. Being inherently interpretable, our threat model allows for a comprehensive analysis and comparison of jailbreak attacks. We find that effective attacks exploit and abuse infrequent bigrams, either selecting the ones absent from real-world text or rare ones, e.g., specific to Reddit or code datasets.