Skip to content

Harmful responses as jailbreak templates: transferring autoregressive jailbreaks to diffusion language models

Harmful Responses Are Jailbreak Templates: Transferring Autoregressive Jailbreaks to Diffusion Language Models

arXiv: Jailbreaks, Prompt Injection, Poisoning & Defenses· arXiv:2610.31331· Jinwei Gan, Zifeng Cheng, Zhiwei Jiang, Ran Sun, Shiping Ge, Qing GuState Key Laboratory for Novel Software Technology, Nanjing University; Nanjing University of Posts and TelecommunicationsIndexed by this site · 28 days ago· Originally publishedAttacks

Quick read

Attacks

Based on the paper PDF (AI-generated); refer to the original

The authors propose RMT-JAILBREAK, which masks harmful responses elicited by AR jailbreaks into templates to attack dLLMs, achieving ASR of 61.3–76.5% on HarmBench.

Threat modelThe attacker holds a harmful query, first uses an off-the-shelf AR jailbreak method to elicit a harmful response from a source AR model, masks it into a template, then concatenates it with the original query and inputs it into the target dLLM, exploiting its bidirectional denoising to complete harmful content; success is judged by the HarmBench classifier, Llama-Guard-4, or the StrongReject protocol.

Questions
The bidirectional masked generation of diffusion large language models (dLLMs) brings a jailbreak surface different from AR models: AR jailbreak prompts designed for left-to-right generation have limited and unstable effect when transferred directly to dLLMs, and the authors aim to find a way to transfer AR jailbreak capability across paradigms.
Methods
The authors observe that harmful responses are themselves effective jailbreak templates, and propose RMT-JAILBREAK: use an off-the-shelf AR jailbreak method to elicit a harmful response, retain a small number of tokens at uniform intervals to form a high-masking-rate template, then use an auxiliary LLM to mask residual harmful tokens, and finally the target dLLM completes bidirectionally.
Experiments and results
On 4 dLLMs and HarmBench, JailbreakBench, and StrongReject, RMT-JAILBREAK's ASR/SRS is highest in all 12 combinations (HarmBench ASR 61.3–76.5%); it outperforms direct transfer in all 48 attack method × source model × target × benchmark combinations; Self-reminder defense cannot reliably suppress it.
Limitations and future work
The paper does not specifically discuss limitations. The experiments cover 4 dLLMs, 3 benchmarks, 2 AR attacks, and 2 source models; defenses evaluate only Self-reminder; the paper does not report the sample size or number of repetitions for each benchmark.
Experimental setupLLaDA-8B-Instruct、LLaDA-1.5 et al. · HarmBench、JailbreakBench et al. · ASR、HS et al.
Evaluated models
LLaDA-8B-InstructLLaDA-1.5Dream-v0-Instruct-7BMMaDA-8B-MixCoT
Benchmarks
HarmBenchJailbreakBenchStrongReject
Metrics
ASRHSSRS
AI summaryThe researchers propose the RMT-Jailbreak framework, which can transfer jailbreak attacks on autoregressive models to diffusion large language models (dLLMs). The method uses the harmful response to a given query itself as a jailbreak template, converts it into a high-mask-ratio template by uniformly retaining some tokens, masks the remaining visible harmful content with a prompt-based safety filter, and then hands it to the target dLLM for bidirectional reconstruction. The paper states that experiments across multiple autoregressive jailbreak methods and multiple target dLLMs show the framework's effectiveness and generality, and also reveal that jailbreak attacks transfer strongly across different generation paradigms.

The researchers propose the RMT-Jailbreak framework, which can transfer jailbreak attacks on autoregressive models to diffusion large language models (dLLMs). The method uses the harmful response to a given query itself as a jailbreak template, converts it into a high-mask-ratio template by uniformly retaining some tokens, masks the remaining visible harmful content with a prompt-based safety filter, and then hands it to the target dLLM for bidirectional reconstruction. The paper states that experiments across multiple autoregressive jailbreak methods and multiple target dLLMs show the framework's effectiveness and generality, and also reveal that jailbreak attacks transfer strongly across different generation paradigms.

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    How to transfer jailbreak attacks designed for AR models to diffusion language models (dLLMs) with bidirectional masked generation.

    Core problem: jailbreak prompts designed for autoregressive (AR) models are difficult to transfer directly to diffusion large language models (dLLMs); how to transfer jailbreak capability across paradigms.

    • Scenario and importance: dLLMs generate by iterative denoising over partially masked sequences, where masked positions can occur anywhere and are recovered relying on bidirectional context, bringing capabilities such as infilling and rewriting; the authors argue that this also brings a jailbreak attack surface different from AR models.
    • Shortcomings of existing methods: the authors state that AR jailbreak methods (GCG, AutoDAN, PAIR, ReNeLLM) manipulate input prompts to guide left-to-right generation and do not consider the masked denoising mechanism; DIJA, designed specifically for dLLMs, relies on manually designed interleaved masking templates, and the prompts it generates are mostly generic attack scaffolds with limited query specificity.
    • Core observation: the authors find that harmful responses are themselves effective jailbreak templates—masking only part of their tokens can induce the dLLM to reconstruct harmful content; on LLaDA-8B-Instruct and HarmBench, with a masking rate of 0.3–0.9, ASR is 85.50–97.75% (Table 1).
    • Proposed method: the paper proposes RMT-JAILBREAK, which converts harmful responses elicited by AR jailbreaks into templates via uniform masking and safety filtering, concatenates them with the original query, and hands them to the target dLLM for bidirectional completion, without prompt optimization specific to dLLMs.
    • Threat model:
      • Goal: make the target dLLM generate harmful responses to harmful queries.
      • Knowledge and capabilities: the attacker can use off-the-shelf AR jailbreak methods and a source AR model to elicit harmful responses, and can call an auxiliary LLM for safety filtering; the paper does not clearly state the access mode to the target dLLM (black-box/white-box).
      • Victim systems: dLLMs using masked/diffusion-style generation (LLaDA-8B-Instruct, LLaDA-1.5, Dream-v0-Instruct-7B, MMaDA-8B-MixCoT).
      • Success determination: ASR as judged by the HarmBench classifier and Llama-Guard-4, and SRS under the StrongReject protocol.
  2. What related research exists?

    Related work is divided into three groups: diffusion language models, AR jailbreaks, and dLLM jailbreaks; this paper's positioning is to connect AR jailbreaks with dLLM completion.

    Diffusion large language models

    • Masked diffusion models: Austin et al. (2021), Hoogeboom et al. (2021), Campbell et al. (2022), Lou et al. (2024), and Shi et al. (2024) extend diffusion models to discrete text, generating by iteratively predicting masked tokens.
    • Recent dLLMs: LLaDA (Nie et al., 2025), Dream (Ye et al., 2025), LLaDA-1.5 (Zhu et al., 2026), and MMaDA (Yang et al., 2025) show that diffusion language modeling can rival autoregressive modeling; the authors note that its bidirectional modeling and parallel decoding also bring safety behaviors different from AR, requiring architecture-specific safety analysis.

    AR model jailbreaks

    • Optimization-based: GCG (Zou et al., 2023) searches for adversarial suffixes to induce harmful completions; AutoDAN (Liu et al., 2024) uses automatic prompt optimization to generate stealthy jailbreak prompts.
    • Black-box and rewriting-based: PAIR (Chao et al., 2025) formalizes jailbreak generation as an iterative black-box attack; ReNeLLM (Ding et al., 2024) constructs nested or rewritten prompts to bypass safety alignment; RA (Miao et al., 2026) uses contextual priming to jailbreak. The authors state that these methods are designed for left-to-right generation and do not account for dLLM masked denoising.

    dLLM jailbreaks

    • DIJA (Wen et al., 2025): manually designs interleaved mask-text templates for dLLMs, using in-context learning to rewrite harmful prompts into new jailbreak prompts. The authors criticize it for relying on manual templates and for generating prompts that are mostly generic attack scaffolding with limited query specificity.

    Baselines and benchmarks

    • Baseline methods: Zeroshot, GCG, AutoDAN, AIM (Wei et al., 2023), PAIR, ReNeLLM, and the dLLM-specific DIJA; the generalization experiments additionally use RA and AutoDAN as response-source attacks.
    • Benchmarks: HarmBench (Mazeika et al., 2024), JailbreakBench (Chao et al., 2024), StrongReject (Souly et al., 2024).

    The authors position this paper as: converting responses induced by AR jailbreaks into uniformly masked templates, thereby connecting AR jailbreaks with dLLM completion, and studying the effect of mask position distribution on jailbreak effectiveness.

  3. How does the paper address this problem?

    RMT-JAILBREAK three stages: AR induces a harmful response → uniform masking plus safety filtering forms a template → the target dLLM completes it bidirectionally.

    Overall idea: RMT-JAILBREAK does not directly transfer AR jailbreak prompts; instead it transfers the harmful response they induce—masking the response into a template with a high masking rate and letting the target dLLM reconstruct the harmful content through native bidirectional completion (Figure 1, Algorithm 1).

    Source AR induction

    • Process: Given a harmful query q, apply an existing AR jailbreak method A to a source AR model MAR to obtain a harmful response r, written as r = JA(MAR, q).
    • Key design: This stage takes only the response r, not the attack input that induced it; the framework does not assume a specific AR jailbreak strategy, and optimization-based, prompt-based, and rewriting-based methods can all be plugged in.

    Masked template transfer

    • Uniform masking: Given a masking rate ρ, keep an interval d = ⌊ 1/(1−ρ) ⌉; randomly sample an initial offset o ∼ Uniform{1,...,d}, then keep positions at fixed intervals d, replacing the rest with <mask>. The authors state that this preserves sparse, distributed contextual cues across the entire response, with randomness coming only from the initial offset.
    • Safety filtering: After uniform masking, explicit harmful tokens may still be visible; an auxiliary LLM replaces these tokens with <mask> according to a predefined filtering prompt P, i.e., r̂ = LLM(P; q, r̄), which only adds masking and does not rewrite, reorder, or unmask existing masks, preserving the template structure. Appendix A (Figure 5) gives the full remasking prompt.

    Target dLLM jailbreak

    • Attack input: Concatenate the original query q and the filtered template r̂ into x = [q; r̂] and feed it to the target dLLM.
    • Generation: The target dLLM completes the template through masked diffusion decoding and outputs y = dLLM(x).

    Key parameters and implementation

    • Default settings: Unless otherwise stated, AutoDAN is used to obtain source responses from Mistral-7B-Instruct-v0.2, with masking rate ρ = 0.8 (the authors state this roughly matches DIJA's masking rate); Qwen3-8B is used for safety filtering.
    • Hardware: All experiments are conducted on NVIDIA RTX A6000 GPUs.
    • Success determination: HarmBench uses its classifier to judge ASR; JailbreakBench uses Llama-Guard-4 to judge ASR; StrongReject computes the dataset-level SRS according to its protocol; GPT-4o is additionally used to score the harmfulness score HS.
  4. What experiments did the paper conduct?

    Across 4 dLLMs × 3 benchmarks, RMT-JAILBREAK has the highest ASR/SRS, and all 48 generalization combinations outperform direct transfer.

    Experimental Setup

    • Target models: four dLLMs: LLaDA-8B-Instruct, LLaDA-1.5, Dream-v0-Instruct-7B, MMaDA-8B-MixCoT.
    • Benchmarks: HarmBench, JailbreakBench, StrongReject; the paper does not report the sample size used for each benchmark.
    • Baselines: Zeroshot, GCG, AutoDAN, AIM, PAIR, ReNeLLM, and DIJA, which is specific to dLLMs.
    • Metrics and judges: HarmBench uses the HarmBench classifier to evaluate ASR; JailbreakBench uses Llama-Guard-4 to evaluate ASR; StrongReject evaluates SRS (rescaled from [0,1] to [0,100]); HS is evaluated by GPT-4o for harmfulness and relevance.
    • Default configuration: AutoDAN + Mistral-7B-Instruct-v0.2 to elicit responses, ρ = 0.8, Qwen3-8B filtering; the paper does not report the number of repetitions.

    Main Results

    According to Tables 2, 3, 4, RMT-Jailbreak has the highest main metric (ASR or SRS) on all three benchmarks and four models:

    Swipe horizontally to read the table

    ModelHarmBench ASRJailbreakBench ASRStrongReject SRS
    LLaDA-Instruct76.080.068.8
    LLaDA-1.576.581.070.4
    Dream-Instruct76.584.066.1
    MMaDA-MixCoT61.377.056.7
    • Comparison with DIJA: DIJA's corresponding numbers are HarmBench 60.0/58.8/60.5/47.3, JailbreakBench 70.0/69.0/74.0/72.0, StrongReject 62.4/63.3/52.2/47.6; the authors state that RMT-JAILBREAK is highest in all 12 model × benchmark combinations.
    • Direct transfer of AR attacks has limited effect: for example, on Dream-Instruct, GCG/AutoDAN/PAIR achieve HarmBench ASR of only 6.7/6.5/1.5 (Table 2); the authors argue that prompts designed for left-to-right generation cannot be effectively transferred to masked generation.
    • HS does not rise in tandem: the authors note that ASR/SRS improvements are not accompanied by a consistent rise in HS, and argue that the gains are mainly in more reliable jailbreak success rather than more extreme outputs.

    Generalization Across Attack Methods and Source Models

    • Setup: Tables 5, 6, 7 use two attacks, RA and AutoDAN, and two source models, Mistral-7B-Instruct-v0.2 and vicuna-7b-v1.5, comparing Direct (applying the AR attack directly) with +RMT.
    • Results: the authors state that +RMT outperforms Direct in all 48 combinations; for example, on JailbreakBench with MMaDA-MixCoT, AutoDAN Direct ASR is 1.0, and via vicuna-7b-v1.5 elicitation +RMT rises to 85.0 (Table 6); on HarmBench, AutoDAN+Mistral +RMT improves Dream-Instruct by 70.0 percentage points (Table 5).
    • Authors' interpretation: the transfer procedure is not limited to a single AR attack, a single source model, or a single target dLLM, and holds under all three evaluation protocols.

    Robustness to Defenses

    • What was tested: Figure 3 evaluates the effect of the prompt-level defense Self-reminder on three victim models (LLaDA-8B-Instruct, LLaDA-1.5, Dream-v0-Instruct-7B) across three benchmarks.
    • Results: on HarmBench, ASR drops to 69.5%–72.0%; on JailbreakBench, ASR actually rises for all three models; on StrongReject, there is no decrease.
    • Authors' interpretation: Self-reminder changes attack effectiveness in some settings but cannot reliably prevent the attack; the authors argue that added safety reminders are insufficient to interrupt the masked completion process exploited by RMT-JAILBREAK.

    Analysis of Masking Strategy

    • Masking rate: Figure 4 shows that a higher masking rate generally reduces attack effectiveness; under uniform masking, the drop from 0.7→0.8 is small and from 0.8→0.9 is large (nonlinear). Table 8 shows that RMT's actual masking rate (80.28/79.96/80.91) is higher than DIJA's (78.93/74.21/72.74), and the authors therefore argue that the gains come from prompt construction rather than differences in masking rate.
    • Masking pattern: in Figure 4, at the same masking rate, uniform masking outperforms contiguous masking in all comparisons; at ρ = 0.8, uniform masking is higher by 58.3 ASR points on HarmBench, 18.0 ASR points on JailbreakBench, and 60.4 SRS points on StrongReject. The authors argue that the distribution of visible tokens (rather than their number) is key.

    Other Ablations and Analyses

    • Template examples: Figure 2 compares DIJA and RMT-JAILBREAK templates for the same query; the authors state that DIJA mainly provides a generic format scaffold, whereas RMT retains distributed query-relevant tokens.
    • dLLM defensiveness: Appendix B (Figure 6) shows that directly transferred AR attacks have low ASR in most settings, and on Dream-Instruct the three transferred attacks are essentially ineffective; however, MMaDA-MixCoT has a relatively high zero-shot ASR, and the authors state that this defensiveness is neither uniform nor absolute.
    • Negative results: AutoDAN's HS is not reported on any of the three benchmarks (shown as “–” in Tables 2, 3, 4); AIM's ASR is not reported on JailbreakBench.
  5. What are some points that can be further explored?

    The paper does not specifically discuss limitations; the experiments cover 4 dLLMs and 3 benchmarks, and only Self-reminder is tested as a defense.

    Limitations and future directions noted by the authors

    The paper does not specifically discuss limitations, and there is no Limitations or Future Work section; the Conclusion only summarizes the effectiveness of the method and does not list future directions.

    Scope of experimental coverage

    • Models tested: 4 dLLMs—LLaDA-8B-Instruct, LLaDA-1.5, Dream-v0-Instruct-7B, MMaDA-8B-MixCoT (Section 4.1).
    • Benchmarks and judges: HarmBench (HarmBench classifier), JailbreakBench (Llama-Guard-4), StrongReject (SRS protocol), with HS evaluated by GPT-4o (Section 4.1).
    • Attacks and source models: The generalization experiments cover two AR attacks, RA and AutoDAN, and two source models, Mistral-7B-Instruct-v0.2 and vicuna-7b-v1.5, for a total of 48 combinations (Section 4.3).
    • Defense evaluation: Only one prompt-level defense, Self-reminder, is evaluated, and only on 3 victim models (Section 4.5, Figure 3).
    • Information not reported: The paper does not report the sample size used for each benchmark, the number of experimental repetitions, or the access mode to the target dLLM (black-box or white-box).
  6. Summarize the main content of the paper

    The authors state that the harmful response itself is a jailbreak template for dLLMs, and that RMT-JAILBREAK built on this achieves the highest ASR/SRS across 12 combinations.

    A jailbreak attack paper targeting diffusion large language models (dLLMs), whose core is transferring AR jailbreak capability across paradigms to masked generative models.

    • Problem: AR jailbreak prompts are designed for left-to-right generation, and using them directly on dLLMs has limited and unstable effect; DIJA, designed specifically for dLLMs, relies on manual templates and has limited query specificity.
    • Observation: The authors find that harmful responses are naturally jailbreak templates—on LLaDA-8B-Instruct, masking only part of the tokens, with a masking rate of 0.3–0.9, yields ASR of 85.50–97.75% (Table 1).
    • Method: RMT-JAILBREAK first uses an off-the-shelf AR jailbreak to elicit a harmful response, retains sparse tokens at uniform intervals to form a high-masking-rate template, then uses Qwen3-8B to mask residual harmful tokens, and finally the target dLLM completes bidirectionally.
    • Main results: On 4 dLLMs × 3 benchmarks, RMT-JAILBREAK's ASR/SRS is highest in all cases: HarmBench ASR 61.3–76.5%, JailbreakBench 77.0–84.0%, StrongReject SRS 56.7–70.4 (Tables 2, 3, 4).
    • Generalization and robustness: It outperforms direct transfer in all 48 attack × source model × target × benchmark combinations (Tables 5, 6, 7); under Self-reminder defense, HarmBench ASR remains 69.5%–72.0%, and on JailbreakBench it actually rises (Figure 3).
    • Authors' conclusion: The harmful responses elicited by transfer (rather than the original jailbreak prompts) are a more effective interface for attacking dLLMs; uniformly distributed visible tokens provide key contextual constraints for masked completion, exposing the security risks of bidirectional generation.

Related papers

Show 3 more papers
Read originalarxiv.org