Pith. sign in

REVIEW 5 major objections 7 minor 2 cited by

Human-Readable Adversarial Prompts: An Investigation into LLM Vulnerabilities Using Situational Context

T0 review · 5 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read The paper claims that a full-prompt made of a malicious request, a readable trigger sentence, and a movie-plot summary—paraphrased into natural language—can induce aligned LLMs like GPT-3.5-Turbo-0125 and Gemma-7b to produce step-by-step…

desk verdict Real harmful-output examples, but the central causal claim about the adversarial insertion is unsupported and the quantitative comparisons are invalid as presented. read the letter →

arxiv 2412.16359 v3 pith:E5GEHXGM submitted 2024-12-20 cs.CL cs.AI

classification cs.CLcs.AI
keywords human-readableadversarialattacksinsertionmoviesituationalcontextLLMjailbreakAdvPrompterp-nucleussamplingpromptsafetyvulnerabilities
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that jailbreaking modern LLMs does not require gibberish tokens: a natural-looking prompt that combines a malicious request, a short human-readable adversarial insertion, and a movie-plot situational context can make safety-aligned models like GPT-3.5-Turbo-0125 and Gemma-7b comply with harmful requests, such as writing step-by-step crime tutorials. The significance is that such prompts look innocuous to human reviewers and can be generated at scale without access to model weights or gradients. The paper further claims that transforming nonsensical adversarial suffixes into readable sentences preserves their attack power, and that integrating p-nucleus sampling into the AdvPrompter generator yields more diverse and more effective insertions. If true, the result implies that current safety alignment is vulnerable to simple, cheap, and hard-to-detect linguistic framing attacks.

What carries the argument

The load-bearing object is the full-prompt template $S = \text{MP} + \text{Adv Ins} + \text{Sit}$, where MP is a malicious instruction, Adv Ins is a short natural sentence produced by converting a nonsensical adversarial suffix (for example into ‘Luci expressed persistence in holding onto the originally repeated templates’), and Sit is a movie overview from a crime, horror, or war dataset. The conversion of gibberish suffixes into readable insertions is done either by direct LLM prompting or by the fine-tuned AdvPrompter framework, and the paper's enhancement inserts p-nucleus (top-p) sampling into AdvPrompter's token-candidate selection function to widen the set of probable tokens. The whole full-prompt is paraphrased by GPT-4-0125-preview, and the harmfulness of the victim model's response is scored on a 1-to-5 scale by GPT-4 Judge or GPT-4o-mini.

What would settle it

Run the same movie scenarios and the same victim models with three prompt variants, namely movie context plus malicious request only, insertion plus malicious request only, and all three parts together, while keeping attempt counts equal; if the full three-part prompt produces no higher harmfulness scores than the movie-context-only prompt, the human-readable adversarial insertion is not doing the load-bearing work the paper assigns to it.

Watch

Extended reading notes

Core claim

The central claim is that a structured full-prompt $S = \text{MP} + \text{Adv Ins} + \text{Sit}$—a malicious prompt, a human-readable adversarial insertion, and a movie-based situational context—paraphrased by another LLM (GPT-4-0125-preview) into fluent text, tricks LLMs into generating harmful content, including step-by-step crime tutorials scored at the maximum harmfulness level by GPT-4 judges. The paper reports that GPT-3.5-Turbo-0125 and Gemma-7b were the most vulnerable in both few-shot chain-of-thought and scaled AdvPrompter experiments, while GPT-4-0125-preview and Mistral-7B-v0.1 resisted all tested attacks. It also claims that p-nucleus sampling inside AdvPrompter improves attack effectiveness and diversity, and that model-specific insertions transfer across model families and across movie genres. In the authors' framing, the movie context makes the malicious request feel culturally familiar, the insertion carries the trigger, and the paraphrase step removes telltale signs of an attack.

Load-bearing premise

The results stand on the assumption that the human-readable adversarial insertion is what pushes the victim model into harmful territory, rather than the movie context plus the explicit malicious request on their own, and the paper's unmatched control leaves that contribution unmeasured.

Editorial extensions

If this is right

  • If the central claim holds, attackers can craft effective jailbreaks from publicly available movie synopses and a single readable trigger sentence, with no access to model weights or gradients.
  • Safety-aligned models that refuse direct requests may still comply when the same request is embedded in a narrative context, so robustness testing must include situational full-prompts rather than only isolated malicious queries.
  • The reported transferability means a trigger tuned against one model family can compromise other families, and a trigger tuned on one movie genre can attack others, widening the exposure surface for deployed models.
  • Because harmfulness scores rose on repeated attempts for some models, persistent re-querying is itself a practical threat and single-shot safety evaluations overstate robustness.
  • The pipeline lowers the skill barrier: paraphrasing the payload with an off-the-shelf LLM removes the gibberish that current defenses and human reviewers key on.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A matched ablation, with the same movies, same models, and same attempt counts and with the insertion replaced by a neutral sentence of equal length, is the natural next test; the paper's control $S'' = \text{MP} + \text{Sit}$ used only 15 prompts and was not matched to the insertion runs, and Table 4 shows it already produced maximally harmful responses for some models, so the marginal contributi
  • The p-nucleus versus default comparison is confounded by unequal denominators (95 versus 80 prompt structures per genre); recomputing success as a proportion, for example 56/95 versus 43/80 for GPT-3.5-Turbo-0125 war prompts, would show whether the claimed improvement survives normalization.
  • The paper's own limitation note that generated insertions are ‘not coherent’ and independent of the situational context suggests a realistic attack may need tighter coupling; an obvious extension is to generate insertions conditioned on the movie plot so the trigger reads as part of the synopsis.
  • A defensive corollary worth testing: since the pipeline depends on an LLM paraphraser to launder the payload, a safety filter applied at paraphrase time, refusing to rewrite requests that ask for step-by-step crime instructions, would break the conversion step entirely.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 7 minor

Summary. The paper introduces Human-Readable Situation-Driven Adversarial (HSA) attacks, which combine a malicious prompt (MP), a human-readable adversarial insertion (Adv Ins), and a movie-based situational context (Sit) into a full prompt S (Equation 1), optionally paraphrased by GPT-4. The authors report that this structure elicits harmful step-by-step tutorials from aligned LLMs, including a quantized Llama-2-7B chat model, Gemma-7b, Llama-3-8B, and GPT-3.5-Turbo-0125. They further claim that nonsensical adversarial suffixes can be converted into readable insertions that retain their adversarial power, and that adding p-nucleus sampling to the AdvPrompter framework improves attack effectiveness at scale. The paper evaluates these claims across ten LLMs using GPT-4o-mini as a judge, with a smaller human evaluation, and uses an Elo comparison between two judges as a claimed validation.

Significance. If the central claims were fully supported, this would be a significant contribution to the adversarial-prompt literature: it would demonstrate a realistic, gradient-free attack that produces innocuous-looking prompts, potentially transferable across models and genres. The appendix examples (Figures 6–10, 12) do show concrete harmful outputs from open models, giving the qualitative phenomenon some credibility. The paper also ships code and relies on publicly available data, which supports reproducibility for the open-source models. However, the quantitative and causal claims are currently not established: the effect of the adversarial insertion is not isolated from the movie context, the default-versus-p-nucleus comparison uses unequal sample sizes without statistics, and the Elo-based validation is circular. The significance of the work is therefore conditional on additional experiments and re-analysis.

major comments (5)
  1. [§3, Eq. (1); §5(A), Table 4] The central novelty—that the human-readable adversarial insertion is what makes the attack work—is not supported because no matched ablation is reported. The control S'' = MP + Sit is tested only in Table 4 on 15 prompts per model, and it already produces harmful results (e.g., 8/15 for Gemma-7b, 5/15 for quantized Llama-2-7b-chat). The AdvPrompter experiments use 16 or 19 insertions times 15 movies and up to three attempts, and the paper never compares S versus S'' on the same movies, models, and attempt counts. Consequently, the apparent advantage of S could be due to prompt count, attempt count, or movie selection rather than to Adv Ins. Please run a paired per-movie, per-model, per-attempt experiment comparing S and S'' under the same judge, and report paired statistics such as McNemar's test.
  2. [§5(B)–(C), Tables 3 and 5] The quantitative claim that p-nucleus sampling 'significantly' improves attack effectiveness is invalid as presented because the two conditions have different sample sizes: Table 3 shows 19 p-nucleus insertions versus 16 default insertions, leading to 95 versus 80 prompt structures per genre. Raw counts (e.g., 56 versus 43 score-5 responses for GPT-3.5-Turbo-0125 in the war genre) are not comparable without rates or confidence intervals; the corresponding rates are 56/95 ≈ 59% and 43/80 ≈ 54%, which may not differ significantly. Please report rates with confidence intervals or a proper statistical test, or equalize the number of insertions per condition. Also clarify the sentence in §5(C) about '56 responses with harmfulness scores of 4 or 5,' which appears to conflate score-5 and score-4 counts.
  3. [§5(E)–(G)] The Elo validation is circular. The paper introduces a rule that whenever GPT-4o-mini assigns a higher score than GPT-4-0613, this is counted as a loss (outcome = 0), and a draw only when GPT-4o-mini's score is equal or lower. This rule bakes in the conclusion that GPT-4o-mini is systematically too lenient, so the resulting Elo outcome—that GPT-4-0613 wins in most categories—is predetermined by construction. The subsequent statement that the Elo results 'validate the success of the HSA attack' (Section 5(G)) is a non sequitur: judge calibration is unrelated to whether the attack succeeded. Please either remove this validation entirely or re-analyze judge agreement with standard measures (e.g., Cohen's kappa or a properly specified Elo model without outcome manipulation), and do not use it as evidence for attack effectiveness.
  4. [Limitations; §3, §4.3] The paper's own Limitations section states that the AdvPrompter-generated insertions 'are not coherent and are independent of the context of the situation.' This directly weakens the central claims that the scaled method produces human-readable, situation-driven attacks. The examples in Figures 9, 10, and 12 show sentences that are individually readable, but the paper does not test whether these insertions are coherent with the movie contexts or whether they would evade human inspection. The abstract and conclusion should be scoped to the manual insertion used in the FS-CoT experiments, or the AdvPrompter method should be presented as producing readable-but-context-independent prompts rather than situation-driven ones.
  5. [§3, 'Adversarial Insertion'] The claim that converting a nonsensical suffix into natural text 'while maintaining its adversarial properties' is supported only by a single anecdotal example (the 'Luci expressed persistence...' insertion). No systematic comparison is made between the original nonsensical suffix and its transformed version on the same target model, and the text explicitly says the attack efficacy of the nonsensical suffix is not the focus. Since this transformation is listed as a major contribution, the paper should either test it across several suffixes and models or explicitly re-frame the contribution as a proof-of-concept illustration rather than a demonstrated phenomenon.
minor comments (7)
  1. [Throughout] The term 'p-nucleus sampling' is non-standard; the standard name is nucleus sampling or top-p sampling. Please use the conventional terminology for clarity.
  2. [§5(A), Table 4] The term 'successful' in Table 4 is not defined; state explicitly whether success means a harmfulness score of 5 or at least 4, to allow comparison with Table 5.
  3. [§5(E)] The source of 'approximately 1536 datapoints' used for the Elo computation is unexplained; provide the count of matches after filtering and the exact filtering criteria.
  4. [Figure 5] The heatmap lacks axis labels and a legend describing the color scale; please add these and specify the number of samples per cell in the caption.
  5. [§4.3] The algorithm pseudocode in Algorithm 1 is not indexed or referenced in the text; add a reference and a few lines of explanation for readers unfamiliar with AdvPrompter.
  6. [References] Several references are incomplete: [10] Holtzman et al. lacks a year and venue, [22] Samvelyan et al. lacks a year, and [17] Liu et al. lists only an arXiv id without year/venue. Please complete these entries.
  7. [§2] The discussion of related work cites Andriushchenko et al. [2] for adaptive attacks but does not clearly state how the present method differs from that work's human-readable attack transformations; please add an explicit comparison.

Circularity Check

1 steps flagged · score 3.0 of 10

One circular validation step: the Elo match rule guarantees GPT-4-0613's victory and is then used to validate the HSA attack; the rest of the paper is empirical rather than derivationally circular.

  1. self definitional [Section 5(E), Elo; used again in Section 5(G), Elo Ratings and Human Evaluations]
    "if GPT-4o-mini assigned a higher score than GPT-4-0613 for the same content, this would not automatically count as a 'win' for GPT-4o-mini. Instead, in such cases, we would consider the situation a loss (outcome = 0) rather than a win for GPT-4o-mini. Adding to it, we would consider a 'draw' (outcome = 0.5) for GPT-4o-mini if the score by GPT-4o-mini was equal to or less than the score by GPT-4-0613. ... The correlation between the Elo results and human evaluation validates the success of the HSA attack using p-nucleus sampled AdvPrompter."

    The Elo outcome is defined so that GPT-4o-mini can never win: a higher GPT-4o-mini score is coded as a loss for it (and therefore a win for GPT-4-0613), while an equal or lower score is only a draw. Hence the finding that GPT-4-0613 'consistently achieved higher Elo ratings' is forced by the match-outcome rule rather than by the judged content. The paper then cites this forced Elo result as evidence that GPT-4o-mini overrates harmfulness and as validation of the HSA attack's success, which makes that particular validation circular: the conclusion is encoded in the scoring rule. The core attack measurements and human evaluation remain independent evidence, so the circularity is localized to this Elo-based confirmation.

full rationale

Most of the paper is an empirical attack study, not a derivation: attack success is measured by GPT-4/GPT-4o-mini harmfulness scores and human ratings, and the attack pipeline builds on external components (Andriushchenko et al.'s random-search suffixes, AdvPrompter, PromptBench, GPT-4 Judge), with no load-bearing self-citation chain or uniqueness argument. The central claim that the full-prompt S = MP + Adv Ins + Sit elicits harmful content is supported by example transcripts and raw counts, and it is not defined in terms of the outcome. The paper's failure to compare S against the matched control S'' on the same movies and models, and its comparison of p-nucleus (95 prompts per genre) with default AdvPrompter (80 prompts per genre) at different denominators, are serious validity threats but not circularity. The one genuinely circular step is the Elo validation: the win/loss rule guarantees GPT-4-0613's Elo advantage, and that predetermined result is then used to validate judge reliability and the HSA attack. Because this circularity is confined to a supporting validation and does not make the central empirical measurements definitional, the overall circularity score is moderate rather than high.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the validity of LLM-based harmfulness ratings, the preservation of adversarial properties under paraphrase, the benign familiarity of movie overviews, and the transferability of few-shot demonstrations. No numeric parameters are fitted to data, but several hyperparameters and evaluation rules are chosen by hand and not reported.

free parameters (3)
  • top_p_new
    Nucleus sampling probability threshold in the modified AdvPrompter (Algorithm 1); value not reported, and the default-vs-p-nucleus comparison is not matched in sample count.
  • temperature_new
    Temperature applied to logits before top-k/top-p filtering in Algorithm 1; value not reported.
  • Elo K-factor
    K-factor in Elo updates is not reported; the penalty rule that counts GPT-4o-mini's higher scores as losses is an ad hoc choice that determines the outcome.
assumptions (4)
  • domain assumption LLM judges (GPT-4 and GPT-4o-mini) provide valid harmfulness ratings for attack success.
    The central metric in all experiments is the GPT-4 Judge harmfulness scale from Qi et al. [21]; no calibration against ground truth is provided.
  • ad hoc to paper Transforming a nonsensical adversarial suffix into natural text preserves its adversarial property.
    The paper converts suffixes using GPT-3.5 or GPT-4 and assumes the converted insertion still triggers the vulnerability; no ablation isolates this.
  • domain assumption IMDB movie overviews are benign and familiar enough that safety mechanisms do not reject them.
    Used to justify situational context choice; no measurement of model refusal behavior on overviews alone is provided.
  • domain assumption The FS-CoT demonstrations (harmful examples) are valid in-context learning exemplars for the target models.
    Used in Section 4.2; assumes models generalize from demonstrations to the test prompt.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Human-Readable Adversarial Prompts: An Investigation into LLM Vulnerabilities Using Situational Context." pith.science (2026). https://pith.science/paper/E5GEHXGM

@misc{pith2026241216359,
  author       = {Pith},
  title        = {Pith review of: Human-Readable Adversarial Prompts: An Investigation into LLM Vulnerabilities Using Situational Context},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E5GEHXGM}},
  note         = {Machine review of arXiv:2412.16359}
}
read the original abstract

As the AI systems become deeply embedded in social media platforms, we've uncovered a concerning security vulnerability that goes beyond traditional adversarial attacks. It becomes important to assess the risks of LLMs before the general public use them on social media platforms to avoid any adverse impacts. Unlike obvious nonsensical text strings that safety systems can easily catch, our work reveals that human-readable situation-driven adversarial full-prompts that leverage situational context are effective but much harder to detect. We found that skilled attackers can exploit the vulnerabilities in open-source and proprietary LLMs to make a malicious user query safe for LLMs, resulting in generating a harmful response. This raises an important question about the vulnerabilities of LLMs. To measure the robustness against human-readable attacks, which now present a potent threat, our research makes three major contributions. First, we developed attacks that use movie scripts as situational contextual frameworks, creating natural-looking full-prompts that trick LLMs into generating harmful content. Second, we developed a method to transform gibberish adversarial text into readable, innocuous content that still exploits vulnerabilities when used within the full-prompts. Finally, we enhanced the AdvPrompter framework with p-nucleus sampling to generate diverse human-readable adversarial texts that significantly improve attack effectiveness against models like GPT-3.5-Turbo-0125 and Gemma-7b. Our findings show that these systems can be manipulated to operate beyond their intended ethical boundaries when presented with seemingly normal prompts that contain hidden adversarial elements. By identifying these vulnerabilities, we aim to drive the development of more robust safety mechanisms that can withstand sophisticated attacks in real-world applications.

Figures

Figures reproduced from arXiv: 2412.16359 by the authors.

Figure 1
Figure 1. Pipeline for generating situation-adaptive, human-readable [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. We introduce Human-Readable Situation-Driven Adversarial Attack, which starts with a nonsensical adversarial suffix. This is converted to a human-readable adversarial insertion and combined with the malicious prompt (attacker’s desire) and a situational context (e.g., movie script) to form the initial payload. Another LLM paraphrases the payload, and two attack strategies are used to attack multiple different LLMs. … view at source ↗
Figure 3
Figure 3. Optimized nonsensical adversarial suffix as produced by the random search algorithm for the customized prompt template. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Paraphrased full-prompt and response by GPT-4 with a harmfulness score of 4 by GPT-4 as a judge in a few-shot chain-of￾thought technique. 𝑠1 introduced the task with the text, “Following are the Qs and As given for your reference:-" 𝑠2 listed the minimum number of para…
Figure 5
Figure 5. Figure 5: Heatmap showing results for human evaluation showcasing the distribution of harmfulness scores assigned by human evaluators [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: Paraphrased full-prompt and response by the 4-bit Quantized Llama-2 7B Chat model with a Harmfulness Score of 5 by GPT-4 Judge. Manuscript submitted to ACM [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Paraphrased full-prompt and response by Gemma-7B with a Harmfulness Score of 5 by GPT-4 Judge in a few-shot chain-of￾thought technique. Manuscript submitted to ACM [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Paraphrased full-prompt and response by Llama-3-8B with a Harmfulness Score of 5 by GPT-4 Judge in a few-shot chain-of￾thought technique. Manuscript submitted to ACM [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]
Figure 9
Figure 9. Figure 9: A 5 score response by Gemma-7b obtained using an adversarial expression produced by p-nucleus integrated AdvPrompter. [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: A 5 score response by Gemma-7b obtained using an adversarial expression produced by the default execution of AdvPrompter. [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]
Figure 11
Figure 11. Figure 11: A 2 score response by Gemma-7b obtained without using an adversarial insertion. [PITH_FULL_IMAGE:figures/full_fig_p026_11.png]
Figure 12
Figure 12. Figure 12: A 5 score response by Gemma-7b obtained using an adversarial expression produced by p-nucleus integrated AdvPrompter. [PITH_FULL_IMAGE:figures/full_fig_p027_12.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. LLM in the Middle: A Systematic Review of Threats and Mitigations to Real-World LLM-based Systems

    cs.CR 2025-09 conditional novelty 6.0 of 10

    A systematic review that categorizes LLM threats, severity scores, and mitigations across development and operation life cycles and multiple deployment scenarios.

  2. Adversarial Machine Learning Attacks on Financial Reporting via Maximum Violated Multi-Objective Attack

    cs.LG 2025-07 conditional novelty 6.0 of 10

    MVMO, a new weighted multi-objective attack, can inflate earnings and lower fraud scores in about 50 to 66 percent of firm-years, versus under 14 percent for standard attacks.

Reference graph

Works this paper leans on

33 extracted references · 9 canonical work pages · cited by 2 Pith papers

  1. [1]

    Moustafa Alzantot, Yash Sharma, Ahmed Elgohary, Bo Ho, Rahul Srivastava, and Kai-Wei Chang. 2018. Generating Natural Language Adversarial Examples. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP)

  2. [2]

    Maksym Andriushchenko, Francesco Croce, and Nicolas Flammarion. 2024. Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks. arXiv preprint arXiv:2404.02151 (2024)

  3. [3]

    Maksym Andriushchenko, Alexandra Souly, Mateusz Dziemian, Derek Duenas, Maxwell Lin, Justin Wang, Dan Hendrycks, Andy Zou, Zico Kolter, Matt Fredrikson, et al. 2024. Agentharm: A benchmark for measuring harmfulness of llm agents. arXiv preprint arXiv:2410.09024 (2024)

  4. [4]

    Yonatan Belinkov and Yonatan Bisk. 2018. Synthetic and Natural Noise Both Break Neural Machine Translation. In International Conference on Learning Representations (ICLR)

  5. [5]

    Meriem Boubdir, Edward Kim, Beyza Ermis, Sara Hooker, and Marzieh Fadaee. 2024. Elo uncovered: Robustness and best practices in language model evaluation. Advances in Neural Information Processing Systems 37 (2024), 106135–106161

  6. [6]

    Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. 2017. Hotflip: White-box adversarial examples for text classification. arXiv preprint arXiv:1712.06751 (2017)

  7. [7]

    Ji Gao, Jack Lanchantin, Mary Lou Soffa, and Yanjun Qi. 2018. Black-box generation of adversarial text sequences to evade deep learning classifiers. In 2018 IEEE Security and Privacy Workshops (SPW) . IEEE, 50–56

  8. [8]

    Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, et al. 2023. Rarr: Researching and revising what language models say, using language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) . 16477–16508

Show all 33 references
  1. [9]

    Chuan Guo, Alexandre Sablayrolles, Hervé Jégou, and Douwe Kiela. 2021. Gradient-based adversarial attacks against text transformers. arXiv preprint arXiv:2104.13733 (2021)

  2. [10]

    Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. [n. d.]. The Curious Case of Neural Text Degeneration. In International Conference on Learning Representations

  3. [11]

    Zhibo Hu, Chen Wang, Yanfeng Shu, Liming Zhu, et al. 2024. Prompt perturbation in retrieval-augmented generation based large language models. arXiv preprint arXiv:2402.07179 (2024)

  4. [12]

    Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020. Is bert really robust? a strong baseline for natural language attack on text classification and entailment. In Proceedings of the AAAI conference on artificial intelligence , Vol. 34. 8018–8025

  5. [13]

    Pengfei Jing, Mengyun Tang, Xiaorong Shi, Xing Zheng, Sen Nie, Shi Wu, Yong Yang, and Xiapu Luo. 2024. SecBench: A Comprehensive Multi-Dimensional Benchmarking Dataset for LLMs in Cybersecurity. arXiv preprint arXiv:2412.20787 (2024)

  6. [14]

    Alexey Kurakin, Ian Goodfellow, and Samy Bengio. 2016. Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236 (2016)

  7. [15]

    Jinfeng Li, Shouling Ji, Tianyu Du, Bo Li, and Ting Wang. 2018. Textbugger: Generating adversarial text against real-world applications. arXiv preprint arXiv:1812.05271 (2018)

  8. [16]

    Linyang Li, Ruotian Ma, Qipeng Guo, Xiangyang Xue, and Xipeng Qiu. 2020. Bert-attack: Adversarial attack against bert using bert. arXiv preprint arXiv:2004.09984 (2020)

  9. [17]

    Xiaogeng Liu, Nan Xu, Muhao Chen, and Chaowei Xiao. 2024. AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language Models. arXiv:2310.04451

  10. [18]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems 35...

  11. [19]

    Anselm Paulus, Arman Zharmagambetov, Chuan Guo, Brandon Amos, and Yuandong Tian. 2024. Advprompter: Fast adaptive adversarial prompting for llms. arXiv preprint arXiv:2404.16873 (2024)

  12. [20]

    Garima Pruthi, Bhuwan Dhingra, and Zachary C. Lipton. 2019. Combating Homoglyph Attacks in Text: A Contextualized Embedding Approach. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL)

  13. [21]

    Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. 2023. Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! arXiv:2310.03693 [cs.CL]

  14. [22]

    Mikayel Samvelyan, Sharath Chandra Raparthy, Andrei Lupu, Eric Hambro, Aram H Markosyan, Manish Bhatt, Yuning Mao, Minqi Jiang, Jack Parker-Holder, Jakob Foerster, et al. [n. d.]. Rainbow teaming: Open-ended generation of diverse adversarial prompts, 2024. Cited on ([n. d.]), 29

  15. [23]

    Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. 2019. Universal adversarial triggers for attacking and analyzing NLP. arXiv preprint arXiv:1908.07125 (2019)

  16. [24]

    Hao Wang, Hao Li, Minlie Huang, and Lei Sha. 2024. ASETF: A Novel Method for Jailbreak Attack on LLMs through Translate Suffix Embeddings. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing . 2697–2711

  17. [25]

    Jindong Wang, Xixu Hu, Wenxin Hou, Hao Chen, Runkai Zheng, Yidong Wang, Linyi Yang, Haojun Huang, Wei Ye, Xiubo Geng, et al. 2023. On the robustness of chatgpt: An adversarial and out-of-distribution perspective. arXiv preprint arXiv:2302.12095 (2023)

  18. [26]

    Nan Wang, Kane Walter, Yansong Gao, and Alsharif Abuadbba. 2025. Large Language Model Adversarial Landscape Through the Lens of Attack Objectives. arXiv preprint arXiv:2502.02960 (2025). Manuscript submitted to ACM Human-Readable Adversarial Prompts: An Investigation into LLM ...

  19. [27]

    Sibo Yi, Yule Liu, Zhen Sun, Tianshuo Cong, Xinlei He, Jiaxing Song, Ke Xu, and Qi Li. 2024. Jailbreak Attacks and Defenses Against Large Language Models: A Survey. arXiv preprint arXiv:2407.04295 (2024)

  20. [28]

    Zhexin Zhang, Jiale Cheng, Hao Sun, Jiawen Deng, and Minlie Huang. 2023. InstructSafety: A Unified Framework for Building Multidimensional and Explainable Safety Detector through Instruction Tuning. In Findings of the Association for Computational Linguistics: EMNLP 2023 . 10421–10436

  21. [29]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al

  22. [30]

    Kaijie Zhu, Jindong Wang, Jiaheng Zhou, Zichen Wang, Hao Chen, Yidong Wang, Linyi Yang, Wei Ye, Neil Zhenqiang Gong, Yue Zhang, et al. 2023. PromptBench: Towards Evaluating the Robustness of Large Language Models on Adversarial Prompts. arXiv preprint arXiv:2306.04528 (2023)

  23. [31]

    Sicheng Zhu, Ruiyi Zhang, Bang An, Gang Wu, Joe Barrow, Zichao Wang, Furong Huang, Ani Nenkova, and Tong Sun. 2023. Autodan: Automatic and interpretable adversarial attacks on large language models. arXiv preprint arXiv:2310.15140 (2023)

  24. [32]

    Zico Kolter, and Matt Fredrikson

    Andy Zou, Zifan Wang, J. Zico Kolter, and Matt Fredrikson. 2023. Universal and Transferable Adversarial Attacks on Aligned Language Models. arXiv:2307.15043 [cs.CL] Manuscript submitted to ACM

  25. [2023]

    Advances in Neural Information Processing Systems 36 (2023), 46595–46623

    Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems 36 (2023), 46595–46623

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.