Saved in:
| Main Author: | |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2604.08557 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866914465995292672 |
|---|---|
| author | Singh, Arth |
| author_facet | Singh, Arth |
| contents | Safety alignment in diffusion language models (dLLMs) relies on a single load-bearing assumption: that committed tokens are permanent. We show that violating this assumption, by re-masking committed refusal tokens and injecting a short affirmative prefix, achieves 74-82% ASR on HarmBench across all three publicly available safety-tuned dLLMs, rising to 92-98% with a generic 8-token compliance prefix. We call this attack TrajHijack; it is the first trajectory-level attack on dLLMs, requires no gradient computation, and generalizes across SFT and preference-optimized (VRPO) models. Three findings emerge. First, the vulnerability is irreducibly two-component: re-masking alone (4.4%) and prefix alone (5.7%) both fail. Second, gradient optimization via a differentiable Gumbel-softmax chain consistently degrades ASR (41.5% vs. 76.1%), because continuous perturbations push token distributions off-manifold. Third, A2D (the strongest published dLLM defense) is more vulnerable to TrajHijack (89.9%) than the undefended model (76.1%): its silent-refusal training removes the contextual resistance that trajectory-level attacks must overcome, an effect we call the Defense Inversion Effect. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2604_08557 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Re-Mask and Redirect: Exploiting Denoising Irreversibility in Diffusion Language Models Singh, Arth Computation and Language Artificial Intelligence I.2.6 Safety alignment in diffusion language models (dLLMs) relies on a single load-bearing assumption: that committed tokens are permanent. We show that violating this assumption, by re-masking committed refusal tokens and injecting a short affirmative prefix, achieves 74-82% ASR on HarmBench across all three publicly available safety-tuned dLLMs, rising to 92-98% with a generic 8-token compliance prefix. We call this attack TrajHijack; it is the first trajectory-level attack on dLLMs, requires no gradient computation, and generalizes across SFT and preference-optimized (VRPO) models. Three findings emerge. First, the vulnerability is irreducibly two-component: re-masking alone (4.4%) and prefix alone (5.7%) both fail. Second, gradient optimization via a differentiable Gumbel-softmax chain consistently degrades ASR (41.5% vs. 76.1%), because continuous perturbations push token distributions off-manifold. Third, A2D (the strongest published dLLM defense) is more vulnerable to TrajHijack (89.9%) than the undefended model (76.1%): its silent-refusal training removes the contextual resistance that trajectory-level attacks must overcome, an effect we call the Defense Inversion Effect. |
| title | Re-Mask and Redirect: Exploiting Denoising Irreversibility in Diffusion Language Models |
| topic | Computation and Language Artificial Intelligence I.2.6 |
| url | https://arxiv.org/abs/2604.08557 |