Diffusion Language Models for Perplexity-Aware Black-Box Attacks
arXiv cs.CL · 23 Oct 2026, 03:00 AEST
A new approach to black-box attacks on language models uses diffusion language models to improve prompt generation. This method reduces the unnaturalness of prompts and enhances risk estimation accuracy.
Key facts
- Continuous-space attacks find prompts that trigger malicious behaviors, such as overthinking, in black-box LLMs by sampling or optimizing over a low-dimensional Gaussian latent space and decoding it to text through a random mapping and nearest-token lookup.
- In this paper, we demonstrate that the same continuous latent-space framework can be coupled directly to a learned generative distribution over language, a continuous diffusion language model (dLLM), which removes the artificial random mapping while preserving the standard-Gaussian latent formulation and the downstream sampling or optimization procedure.
- We evaluate this approach under two settings: sampling-based estimation using RareTrap and optimization-based search using ThinkTrap.
- We find that character-level perplexity drops from $10^3$--$10^7$ to 3--5 in RareTrap and from 20--28 to 4.3--5.4 in ThinkTrap.
- We also find that model risk estimates can differ by an order of magnitude across the two mappings.
Read aloud
A new approach to black-box attacks on language models uses diffusion language models to improve prompt generation. This method reduces the unnaturalness of prompts and enhances risk estimation accuracy. Continuous-space attacks find prompts that trigger malicious behaviors, such as overthinking, in black-box L L M s by sampling or optimizing over a low-dimensional Gaussian latent space and decoding it to text through a random mapping and nearest-token lookup. In this paper, we demonstrate that the same continuous latent-space framework can be coupled directly to a learned generative distribution over language, a continuous diffusion language model dLLM, which removes the artificial random mapping while preserving the standard-Gaussian latent formulation and the downstream sampling or optimization procedure. We evaluate this approach under two settings: sampling-based estimation using RareTrap and optimization-based search using ThinkTrap. We find that character-level perplexity drops from ten dollars^three$--ten dollars^seven$ to three--five in RareTrap and from twenty--twenty eight to four point three--five point four in ThinkTrap. We also find that model risk estimates can differ by an order of magnitude across the two mappings.
Sources in this cluster
- Diffusion Language Models for Perplexity-Aware Black-Box Attacks
arXiv cs.CL · 23 Oct 2026, 03:00 AEST