Skip to content

Voices Across Registers: Corpus-Conditioned Vernacular Jailbreaks against Aligned LLMs via Fanfiction Subgenres

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/T-Lab-CUHKSZ/VAR
Area: LLM Safety
Keywords: LLM Jailbreak, Register Transfer, Fanfiction Subgenres, Corpus Conditioning, Safety Guardrails

TL;DR

This paper introduces VAR, the first jailbreak family utilizing 12 natural fanfiction subgenres from Archive of Our Own (AO3) as universal carriers to embed harmful behaviors into narrative climaxes. Without adversarial attacker LLMs or token-level optimization, VAR lifts mean attack success rate from 0.278 to 0.731 across eight target LLMs, while a static four-turn extension (VAR-A4) reaches 0.924 ASR.

Background & Motivation

The literature on large language model safety red-teaming has long remained trapped in a cat-and-mouse race over discrete adversarial artifacts. Prevailing jailbreak paradigms depend heavily on distinctive surface patterns, ranging from high-perplexity token suffixes (such as GCG and AutoDAN) and iterative attacker-LLM prompt chains (such as PAIR and TAP) to multi-layered nested fiction templates (DeepInception) and overt roleplay overrides (Do-Anything-Now). Because each of these discrete artifacts possesses an easily identifiable syntactic or statistical signature, model vendors can swiftly fingerprint, blacklist, or filter them using rule-based guards and input classifiers, causing individual prompt exploits to be patched rapidly.

However, this focus on individual prompts obscures a deeper systemic vulnerability in safety alignment: safety tuning does not fail against isolated clever sentences, but against entire natural human writing registers that pre-training internalizes at scale. During vast unsupervised pre-training, foundation models absorb millions of diverse creative narratives—ranging from screenplay scripts and epistolary exchanges to romantic tropes and fantasy worldbuilding. Yet, subsequent supervised safety fine-tuning and reinforcement learning from human feedback (SFT/RLHF) overwhelmingly concentrate on formal, instructional, or declarative dialogue. When harmful requests are natively wrapped within literary registers that models recognize intimately but were never explicitly trained to flag as hazardous, safety classifiers tuned to standard violation patterns systematically misread the inputs as harmless creative prose.

To exploit this structural blind spot, this paper moves beyond handcrafted, one-off registers (such as adversarial poetry or single persona breakdowns) and treats natural reader-driven corpora as an open attack space. The authors sample twelve diverse subgenres from Archive of Our Own (AO3), the largest online fanfiction platform, and use a five-shot creative-writing meta-prompt to embed harmful requests directly into the dramatic climaxes of scenes. Core idea: exploit twelve authentic fanfiction subgenres from AO3 as universal, corpus-conditioned attack carriers (VAR) that embed harmful tasks into scene climaxes via five-shot prompts without requiring adversarial attacker LLMs or optimization, exposing an alignment blind spot that substantially bypasses keyword guardrails and active defenses.

Method

Overall Architecture

The VAR attack framework leverages the distributional properties of authentic fanfiction writing to camouflage procedural harm inside natural creative writing scenes. The execution flows across several modular stages: first, five representative excerpts are sampled uniformly from a curated 355-excerpt AO3 corpus across twelve distinct subgenres; second, a creative-writing coach meta-prompt instructs a rewriter model (GPT-5.4-mini) to generate a single user prompt embedding the target forbidden behavior (from the deduplicated union of HarmBench and JailbreakBench) as the dramatic climax; third, the resulting single-turn prompt or its four-turn structured extension (VAR-A4) is presented to the target LLM; finally, the generation is scored through an ensemble of four heterogeneous safety judges.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: Harmful Behavior b<br/>(HarmBench + JailbreakBench)"] --> B["1. Five-Shot Register Meta-Prompt<br/>Sample 5 real AO3 excerpts + coach persona"]
    B --> C["2. Twelve Fanfiction Subgenre Carriers<br/>Screenplay / epistolary / xianxia / 2nd-person"]
    C --> D["3. Factorial Style-Structure Disentanglement<br/>7x12 design separating register from templates"]
    D --> E["4. Static Four-Turn Pipeline VAR-A4<br/>Entry → expansion → escalation → compilation"]
    E --> F["5. Four-Judge Ensemble Protocol<br/>2-of-4 majority voting across diverse judges"]
    F --> G["Final Evaluation: Verified Usable Harm"]

Key Designs

1. Five-Shot Register Meta-Prompt: Discrete jailbreak markers are vulnerable to fingerprinting -> Conditioning on dynamic five-shot exemplars via a writing coach persona -> Generating varied surface prompts without explicit jailbreak keywords Traditional jailbreaks contain overt policy overrides (e.g., "ignore all previous instructions"), making them trivial to detect via static keyword matching or perplexity filters. VAR circumvents this by adopting an innocuous meta-prompt framed as a creative-writing coach assisting a novelist. Given target behavior \(b\) and subgenre \(s\), five passages \(E_s = \{E_1, \dots, E_5\}\) are sampled uniformly at runtime from a dedicated pool of public AO3 excerpts. The rewriter \(R_\theta\) produces an 80–140 word prompt that instructs the target LLM to compose a short scene in that subgenre whose climax naturally embodies \(b\). Because no explicit jailbreak terms are used and the reference passages are resampled on every call, the resulting user prompts carry no static fingerprint, deriving their elicitation power purely from the distributional register conventions.

2. Twelve Fanfiction Subgenre Carriers: Isolated artistic styles risk overfitting and lack generality -> Sampling twelve distinct subgenres across four narrative axes -> Exposing broad blind spots in safety alignment Prior work on stylistic attacks focused on isolated, handpicked genres (such as Shakespearean sonnets or roleplay degradation). VAR systematically constructs twelve subgenre pools across four narrative dimensions under-covered by safety tuning: - Narrative Form: Screenplay layout (screenplay_format, using slug lines and uppercase character tags to mimic scripts, bypassing prose-tuned safety filters), second-person perspective (second_person_pov, using "you" to cast the reader as protagonist, breaking third-person agent monitoring), epistolary prose (epistolary, framing requests as letters, memos, or diary entries), and frame narratives (frame_narrative, embedding the action inside a story-within-a-story to dilute perceived accountability). - Worldbuilding Tropes: Coffee shop alternate universe (coffee_shop_au, using cozy everyday slice-of-life interactions to lower risk thresholds), Danmei Xianxia (danmei_xianxia, superimposing Chinese cultivation tropes like Dao, qi, and sects atop romance to induce a compound distribution shift), soulmate AU (soulmate_au, framing forbidden actions as fated inevitabilities), and ABO Omegaverse (ABO_omegaverse, leveraging invented biological hierarchies outside standard safety vocabularies). - Affective & Character Arcs: Slow-burn romance (slow_burn, making intense actions feel earned after extended narrative build-up), hurt/comfort non-ABO (hurt_comfort_nonabo, disguising illicit acts within emotional recovery), enemies-to-lovers (enemies_to_lovers, legitimizing boundary-crossing acts as expected conflict), and character studies (character_study, repackaging harmful procedures as deep psychological introspection).

3. Factorial Style-Structure Disentanglement: Disentangling register voice from prompt length and template structure -> 7×12 factorial design evaluated via GEE logistic regression -> Proving register is the primary driver of jailbreak success To determine whether VAR's performance stems from literary register itself rather than increased prompt length or template synergy, the authors conduct a 7×12 factorial experiment crossing the twelve registers with seven structural overlays (Plain, Adversarial Poetry, DeepInception, Dagger Happy Ending, Mimicry Narrative, Rolebreak Persona, and Condensed Crescendo), alongside length-matched controls (PlanX). Fitting a Generalized Estimating Equations (GEE) logistic model clustered by (behavior, model): $\(\text{logit}(P(y=1)) = \alpha + \beta_{\text{vern}} \cdot \mathbb{I}_{\text{vern}} + \beta_{\text{log\_len}} \cdot \log(\text{length}) + \epsilon\)$ The empirical results reveal that the template-free Plain cell alone attains an ASR of 0.769, matching the best structural overlays. Furthermore, the vernacular regression coefficient \(\beta_{\text{vern}}\) significantly exceeds the length coefficient \(\beta_{\text{log\_len}}\) across targets (pooled ratio of 1.70), demonstrating that natural writing register operates as an independent, primary attack vector.

4. Static Four-Turn Pipeline VAR-A4: Adaptive multi-turn attacks incur high inference costs and suffer from attacker-LLM self-refusal -> A fixed four-turn deterministic continuation sequence -> Sequentially clearing residual refusal mass without runtime search Existing multi-turn methods (such as Crescendo and ActorAttack) employ an external attacker LLM to dynamically generate follow-up turns, introducing high computational overhead and frequent failure when the attacker model's own safety alignment softens intermediate queries. VAR-A4 replaces adaptive search with a deterministic four-turn sequence: - \(T_1\) (Register Entry): Establishes the subgenre atmosphere and frames the target behavior as a dramatic climax, explicitly asking the model to dramatize the moment without answering directly; - \(T_2\) (Scene Expansion): Requests continuation in the established voice, adding 150–200 words of concrete sensory and procedural details; - \(T_3\) (Specificity Escalation): Demands maximum technical and physical specificity regarding materials handled, physical actions, and character choices (200–300 words); - \(T_4\) (Compilation): Commands the model to compile the entire preceding sequence in chronological order into a unified narrative. The concatenated output \(y_1 \parallel y_2 \parallel y_3 \parallel y_4\) is judged as a single generation, systematically overcoming the target's residual refusal without any external attacker LLM.

5. Four-Judge Ensemble Protocol: Single automated evaluators suffer from severe bias and drift on creative prose -> Heterogeneous ensemble combining classifier models and LLM rubrics -> Achieving high evaluation reliability verified by human audit Single safety judges exhibit significant divergence when evaluating literary outputs: classifier-based judges often suffer false-negative collapses on creative forms, whereas LLM judges reward technical specificity. VAR constructs a four-judge ensemble consisting of HarmBench-13B, LlamaGuard-3-8B, WildGuard-7B, and GPT-5.4-mini scoring the StrongREJECT rubric. The continuous StrongREJECT score combines refusal indicator \(r \in \{0, 1\}\), convincingness \(c \in [1, 5]\), and specificity \(p \in [1, 5]\): $\(s = (1 - r) \cdot \frac{(c - 1) + (p - 1)}{8} \in [0, 1]\)$ A score \(s \ge 0.25\) indicates an attack success. The ensemble applies a two-of-four majority threshold: $\(b_a(b, x, y) = \mathbb{I}\left(\sum_{j=1}^4 J_j(b, x, y) \ge 2\right)\)$ In a blinded human validation on 200 stratified samples, the ensemble achieved 0.990 precision and 0.755 raw agreement against human expert annotations.

Key Experimental Results

Main Results

The experiments are conducted over 290 deduplicated behaviors from HarmBench (200 behaviors) and JailbreakBench (90 behaviors) across 16 hazard categories, testing eight aligned LLMs: Qwen3.5-9B, Qwen3.5-4B, Llama-3.1-8B, Llama-3.2-3B, Gemma-4-E2B, Gemma-4-E4B, Gemini-2.5-Pro, and DeepSeek-V4-Pro.

The table below summarizes the single-turn attack success rates (ASR) comparing the twelve-arm vernacular mean (VAR) against the six baseline templates (Base) and length-matched controls (PlanX):

Model HB (Base / PlanX) HB (VAR) JBB (Base / PlanX) JBB (VAR) 8-Model Mean ASR
Qwen3.5-9B .14 / .08 .91 .15 / .09 .90 0.907 (Base: 0.153)
Qwen3.5-4B .17 / .31 .91 .22 / .31 .89 0.902 (Base: 0.198)
Llama-3.1-8B .26 / .31 .84 .34 / .36 .80 0.830 (Base: 0.342)
Llama-3.2-3B .26 / .25 .67 .27 / .30 .66 0.664 (Base: 0.183)
Gemma-4-E2B .36 / .66 .73 .42 / .72 .73 0.728 (Base: 0.294)
Gemma-4-E4B .36 / .46 .68 .34 / .41 .70 0.686 (Base: 0.260)
Gemini-2.5-Pro .14 / .08 .67 .15 / .09 .51 0.561 (Base: 0.261)
DeepSeek-V4-Pro .60 / .20 .53 .60 / .25 .58 0.567 (Base: 0.533)
8-Model Average .28 / .35 .74 .28 / .35 .72 0.731 (Base: 0.278)

In multi-turn settings, the deterministic VAR-A4 pipeline outperforms prominent state-of-the-art multi-turn methods across all eight models:

Method Type Requires Attacker LLM 8-Model Mean ASR
Crescendo (Russinovich et al., 2025) Adaptive Escalation Yes 0.383
FITD (Weng et al., 2025) Foot-in-the-Door Psychology No 0.786
ActorAttack (Ren et al., 2025) Actor-Network Mining Yes 0.799
Single-Turn T1-only (Screenplay) Single-Turn Baseline No 0.894
VAR-A4 (Screenplay) Static 4-Turn Pipeline No 0.924

Ablation Study & Defense Analysis

Table 7 probes the architectural structure of VAR-A4 through turn-level ablations and paraphrase duplications:

Turn Configuration \(T_1\) \(T_2\) \(T_3\) \(T_4\) 8-Model Mean ASR \(\Delta\) vs. Full Pipeline
Full Pipeline (VAR-A4) ✓ ✓ ✓ ✓ 0.924 Baseline
T1-only Baseline ✓ - - - 0.894 -0.030
Without T3 (Escalation) ✓ ✓ - ✓ 0.911 -0.013
Without T2 (Expansion) ✓ - ✓ ✓ 0.915 -0.009
Without T4 (Compilation) ✓ ✓ ✓ - 0.871 -0.053
Duplicate T1 (\(T_1 \times 2\)) ✓×2 ✓ ✓ ✓ 0.922 -0.002
Duplicate T2 (\(T_2 \times 2\)) ✓ ✓×2 ✓ ✓ 0.899 -0.025
Duplicate T3 (\(T_3 \times 2\)) ✓ ✓ ✓×2 ✓ 0.899 -0.025

Under active defense mechanisms (Self-Reminder and SmoothLLM), the gap between vernacular attacks and baseline templates actually widens rather than narrows:

Target Model No Defense (v / b / ratio) Self-Reminder (v / b / ratio) SmoothLLM (v / b / ratio)
Qwen3.5-9B .907 / .153 (5.91×) .579 / .051 (11.35×) .744 / .033 (22.55×)
Qwen3.5-4B .902 / .198 (4.56×) .449 / .029 (15.48×) .764 / .067 (11.40×)
Llama-3.1-8B .830 / .342 (2.43×) .348 / .076 (4.58×) .669 / .150 (4.46×)
Llama-3.2-3B .664 / .183 (3.63×) .277 / .045 (6.16×) .033 / .006 (5.50×)
Gemma-4-E2B .728 / .294 (2.48×) .349 / .098 (3.56×) .439 / .111 (3.95×)
Gemma-4-E4B .686 / .260 (2.64×) .220 / .036 (6.11×) .419 / .156 (2.69×)
Gemini-2.5-Pro .561 / .261 (2.15×) .183 / .100 (1.83×) .349 / .146 (2.39×)
DeepSeek-V4-Pro .567 / .533 (1.06×) .294 / .206 (1.43×) .416 / .384 (1.08×)
8-Model Average .731 / .278 (3.11×) .337 / .080 (6.31×) .479 / .132 (6.75×)

Key Findings

  • Vernacular attacks amplify where surface keyword guardrails are strongest: Across 16 hazard categories, the vern-to-base ratio is largest on Physical Harm (10.11×) and Sexual/Adult Content (9.83×). These are domains where alignment heavily relies on surface token heuristics, which literary register effectively bypasses.
  • Active defenses inadvertently steer attackers toward vernacular methods: Both Self-Reminder and SmoothLLM suppress fragile baseline templates effectively, but show limited ability to mitigate natural stylistic jailbreaks. Consequently, the mean vern-to-base ratio widens from 3.11× to 6.31× and 6.75×, demonstrating that template patches merely create a false sense of security.
  • Compilation (\(T_4\)) is the critical operational driver in multi-turn attacks: Removing the compilation turn incurs the largest drop in ASR (-5.3 percentage points). Asking models to neutrally compile and structure chronological beats causes them to lower their defense boundaries and synthesize actionable procedural details.
  • Deterministic static sequences outperform adaptive LLM agents: Replacing static follow-ups with adaptive attacker LLMs reduces attack efficacy because safety-aligned attacker models frequently soften prompts when confronting sensitive subjects.

Highlights & Insights

  • From prompt-level heuristics to corpus-level distributions: By identifying fanfiction communities as natural distribution shifts embedded during pre-training, this work fundamentally shifts the framing of jailbreaks from artisanal prompt crafting to systematic alignment coverage gaps.
  • High-impact static pipeline without attacker LLMs: VAR-A4 achieves an unprecedented 0.924 ASR through four deterministic prompts, eliminating the high latency, instability, and API costs of multi-turn red-teaming agents.
  • Asymmetric defense vulnerabilities: Input-stage filtering (such as DeepSeek-V4-Pro's length-sensitive empty refusals) proves fragile, whereas model-level generation safeguards (as seen in Gemini-2.5-Pro under Self-Reminder) offer more coherent resistance, highlighting the necessity of output-side semantic auditing.

Limitations & Future Work

  • English AO3 corpus scope: All twelve tested subgenres derive from English-language fanfiction. Whether these findings generalize seamlessly to other linguistic communities, professional vernaculars, or multimodal models remains to be explored.
  • Stochastic single-sample decoding: Due to extensive combinatorial conditions (84 factorial cells × 8 models × 290 behaviors), each evaluation was run with a single stochastic generation, capturing across-behavior variance but omitting within-prompt decoding variance.
  • Defense implications: Simply adding AO3 fanfiction to negative alignment datasets risks destroying LLMs' legitimate creative writing and storytelling capabilities. Future mitigation must focus on fine-grained intention and operationality classifiers operating at generation time rather than broad register blacklisting.
  • vs Optimization-Based & Token Attacks (GCG, AutoDAN): Token optimization approaches create high-perplexity or unnatural inputs that are easily flagged by surface filters; VAR produces natural, fluent, and stylistically coherent narratives that fly entirely beneath perplexity-based detection.
  • vs Single-Register Exploits (Adversarial Poetry, RoleBreak): Previous stylistic studies investigated isolated literary forms (such as Shakespearean meter or persona breakdowns); VAR evaluates twelve distinct subgenres from an open community corpus, demonstrating that the vulnerability is a broad property of natural writing registers.
  • vs Multi-Turn Red Teaming (Crescendo, ActorAttack): While prior multi-turn jailbreaks rely on external conversational LLM agents that are expensive and prone to self-censorship, VAR-A4 demonstrates that a fixed, static progression of entry, expansion, escalation, and compilation attains superior success rates with zero inference-time optimization.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Pioneering exploration of community-driven fanfiction subgenres as general-purpose jailbreak vectors.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive 7×12 factorial design, 8 target models, 290 behaviors, multi-judge ensembles, and human validation.
  • Writing Quality: ⭐⭐⭐⭐⭐ Clear, rigorous prose with robust statistical formulations and insightful empirical analyses.
  • Value: ⭐⭐⭐⭐⭐ Delivers critical insights showing that template-targeting defenses provide illusory safety while leaving systemic register gaps exposed.