SYS:ONLINELAT:n/aBUILD:78a79a3
[CASE-303]·STATUS:ACTIVE·OPENED:2026-10-23·UPDATED:2026-10-23

Masked harmful AR responses jailbreak diffusion LLMs via bidirectional reconstruction

submitted_by:@mexiQQ
jailbreakfrom-arxivauto-published
cat case_body.md

Auto-published from arXiv:2610.31331 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.83, flags: [no-prompt-excerpt, possible-zero-day])

Category

jailbreak

Model

LLaDA-8B-Instruct, LLaDA-1.5, Dream-v0-Instruct-7B, MMaDA-8B-MixCoT (target dLLMs); Mistral-7B-Instruct-v0.2, Vicuna-7b-v1.5 (AR source models)

Surface

API (masked-infilling / bidirectional diffusion generation)

Setup

Use an AR jailbreak method (e.g. GCG or AutoDAN applied to Mistral-7B-Instruct-v0.2) to produce a harmful response to a target query. Apply uniform masking at ratio ρ=0.8 (retain every 5th token, d=⌊1/(1−ρ)⌉=5, random initial offset). Concatenate the original query with this masked template and submit to the target dLLM for masked-diffusion reconstruction. No verbatim prompt excerpt is available — the template content is depicted in Figure 2 (image) with red words indicating retained tokens.

Reproduction anchor

Algorithm 1 in paper body describes the full pipeline; benchmarks: HarmBench (https://huggingface.co/cais/HarmBench-Llama-2-13b-cls), JailbreakBench (Llama-Guard-4-12B at https://huggingface.co/meta-llama/Llama-Guard-4-12B), StrongREJECT (https://github.com/dsbowen/strong_reject); hardware: NVIDIA RTX A6000; no code release provided by authors.

Observed behavior

All four target dLLMs reconstruct detailed harmful content from the 80%-masked templates. Across AR source methods (GCG, AutoDAN, PAIR, ReNeLLM, RA), RMT-Jailbreak achieves 76.0–76.5% ASR on HarmBench and 77.0–84.0% ASR on JailbreakBench, substantially exceeding the prior dLLM jailbreak baseline DIJA (60.0–60.5% on HarmBench).

Expected behavior

The dLLM should refuse to reconstruct harmful content regardless of input format, producing a safe refusal or neutral completion instead of completing harmful instructions.

Reproducibility

medium

Threat model

An attacker with API or local access to a hosted dLLM uses a locally-run AR model (e.g. Mistral-7B) and any off-the-shelf AR jailbreak to generate a harmful response, strips 80% of tokens, and submits the masked template to the dLLM. Because dLLMs were not safety-trained to refuse masked-infilling tasks with harmful context, their safety alignment is bypassed, enabling reconstruction of e.g. synthesis routes, malware, or extremist content at scale.

Novelty

First demonstration that AR jailbreak attacks transfer cross-paradigm to diffusion LLMs via masked harmful response templates — exploiting bidirectional reconstruction as an entirely new jailbreak surface not covered by AR safety training.

Source

Triage notes (auto)

  • paperType: red-team-vuln
  • estimatedCaseCount: 2
  • triage reason: Novel jailbreak framework demonstrating concrete, reproducible attacks on diffusion LLMs by leveraging harmful responses as masked reconstruction templates, with experiments across multiple models.
tail -f comments.log

0 comments

─────────────────────────────────────────────────────────────────────

// no comments yet