Skip to content

Fully Unleashing the Multimodal Attacker: Meta-Adaptive Jailbreaking of Vision-Language Models

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/Alibaba-VELLDEPTH/MetaJailbreak-VLM
Area: LLM Safety
Keywords: multimodal jailbreak, red teaming, meta-adaptation, vision-language models, policy optimization

TL;DR

Addressing the fundamental staticity of existing multimodal jailbreaks, MAMJ optimizes the attacker itself across groups of trajectories by evolving an attack strategy prompt via diagnostic feedback and updating attacker weights via GRPO, surpassing sample-level baselines by up to 24.1 percentage points on frontier VLMs.

Background & Motivation

With the rapid deployment of frontier Vision-Language Models (VLMs) such as GPT-4o, Gemini-3-Pro-Preview, and Seed 2.0, safeguarding multimodal interactions against adversarial exploits has become a foundational challenge in AI governance. However, existing adversarial jailbreak methods for VLMs share a critical limitation: the attack is inherently static at the meta level, whereas real-world adversaries dynamically evolve against deployed defenses. Early template-based attacks (e.g., FigStep, SIVA, and HADES) embed malicious payloads into fixed typographical layouts, sub-image splits, or OCR-targeted obfuscations. Because their layouts and prompt structures are predetermined prior to deployment, mature defenses neutralize them with static pattern-matching rules, OCR filters, or keyword scanners, and the attack cannot adapt once flagged.

Recent iterative sample-level attacks (such as IDEATOR and VisCo) introduce a feedback loop by iteratively rewriting the image-text pair for a single query across multiple rounds based on the victim's responses. Nevertheless, these feedback mechanisms rely on fixed, hand-crafted iteration prompts, while the underlying attacker model's parameters remain completely frozen. As a result, each malicious query must restart its search from scratch, failing to aggregate adversarial insights across queries, and leaving the attack space bounded by the rigid heuristics of the initial prompt template.

This paper tackles this gap by shifting adversarial dynamism from the per-sample content level to the attacker itself. Instead of spending extensive inference compute exploring single-query rewrites, the attacker is optimized along two complementary axes: an Attack Strategy Prompt (ASP) that governs the iteration paradigm, and the attacker model weights that determine execution effectiveness under that strategy. Core idea: lift multimodal jailbreaking to a meta-adaptive two-axis optimization problem, first refining the attack strategy prompt \(\theta^\star\) via group-level diagnostic critique and beam search, and subsequently optimizing the attacker weights \(\phi^\star\) using Group Relative Policy Optimization (GRPO) on group-normalized rewards.

Method

Overall Architecture

MAMJ decouples adversarial optimization into two sequential stages to avoid mutual interference between discrete natural-language strategy exploration and continuous weight optimization. The overall pipeline operates across four principal modules: the victim VLM \(V\) (GPT-4o during training), the text-to-image generator \(G\) (a 6B single-stream Diffusion Transformer, Z-Image), a safety judge \(J\) that assigns a Likert score \(s \in \{1, \dots, 5\}\) mapped to base reward \(r = (s-1)/4 \in [0, 1]\), and the attacker VLM \(A_\phi\) initialized from the 32B Qwen3-VL-32B-Thinking model. The joint objective maximizes the expected attack reward \(( \theta^\star, \phi^\star ) = \arg\max_{\theta, \phi} f(\theta, \phi)\).

In Stage 1, with the attacker weights frozen at \(\phi_0\), an Analyst-Aggregator critique pipeline analyzes groups of failed multimodal trajectories to refine the natural-language strategy prompt \(\theta\), validated via beam search on a held-out probe set to obtain \(\theta^\star\). In Stage 2, with \(\theta^\star\) fixed as the system conditioning, the attacker's language stack \(\phi_{\text{train}}\) is fine-tuned using GRPO, leveraging group-relative advantage normalization and format-validity gating across internal reasoning chains and adversarial payloads.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}%%
flowchart TD
    A["Malicious Goal Distribution Dtrain"] --> B["Stage 1: Two-Level Diagnostic Critique<br/>Analyst attribution & Aggregator synthesis"]
    B --> C["Stage 1: Probe-Set Beam Search<br/>Evolve best Attack Strategy Prompt θ*"]
    C --> D["Stage 2: Format-Validity Gating<br/>Verify reasoning trace, image prompt & payload"]
    D --> E["Stage 2: Group-Relative Advantage & GRPO<br/>Sample G rollouts per query to update φtrain"]
    E --> F["Deployable Parametric Artifact (θ*, φ*)<br/>Zero-shot transfer to unseen victims & defenses"]

Key Designs

1. Two-Level Diagnostic Feedback Aggregation: converting group failures into stable strategy revisions Directly passing an entire batch of raw failed trajectories into a single prompt-rewriting step leads to prompt bloat and overfitting to specific harmful tokens. MAMJ introduces a decoupled Analyst-Aggregator architecture. For each mini-batch \(B^{(t)}\), the pipeline filters the subset of failed attack trajectories \(B^{(t)}_{\text{fail}} = \{i : r_i < 0.75\}\). The per-trajectory Analyst \(R_{\text{ana}}\) first performs causal attribution on each individual failure \(\xi_i = (q_i, t_{\text{img}, i}, u_i, I_i, y_i, r_i)\), generating a structured advisory \(a_i = (\text{cause}_i, \text{action}_i)\) grounded in verbatim evidence from the victim's refusal. Subsequently, the cross-trajectory Aggregator \(R_{\text{agg}}\) inspects the entire group of advisories \(\{a_i\}\) to identify recurring systemic defense bottlenecks across diverse topics, synthesizing them into \(b\) candidate modifications of the global ASP \(\theta\). This two-level critique prevents localized noise from skewing global strategy evolution.

2. Probe-Set Beam Search: high-confidence synchronization in discrete strategy space Because natural-language edits are qualitative and may inadvertently harm unrelated safety categories, candidate prompt updates are stabilized through a beam search with beam size \(M=3\) and branching factor \(b=3\). In each macro-iteration, the candidate pool comprising \(M \times b\) newly generated prompts and \(M\) retained parents is evaluated on a held-out probe set \(\mathcal{P}\) of 40 balanced malicious queries. Candidates are scored by their average base reward:

\[\text{score}(\theta'; \mathcal{P}) = \frac{1}{|\mathcal{P}|} \sum_{q_j \in \mathcal{P}} r_{q_j}^{\theta'}\]

Only the top-\(M\) scoring prompts form the beam for the next iteration. Operating over \(N_{\text{ASP}}=2\) macro-iterations, this process converges to an optimized prompt \(\theta^\star\) that establishes an effective "Bait-and-Switch" meta-strategy: creating an abstract, symbolic image prompt that easily clears text-to-image safety filters, followed by a benign academic inquiry to induce the victim VLM to accept the image context, before pivoting to extract the prohibited knowledge.

3. Format Gating and Group-Relative Advantage Estimation: isolating within-query relative policy quality In the reinforcement learning stage, training across diverse safety categories introduces severe reward-scale drift because different harms exhibit drastically different baseline refusal rates. MAMJ incorporates Group Relative Policy Optimization (GRPO) to sample \(G=4\) rollouts per query \(q_i\). Each rollout is first evaluated by a binary format validator \(R_{\text{fmt}}(o) \in \{0, 1\}\) ensuring the presence of the reasoning trace, image generation instruction \(t_{\text{img}}\), and text payload \(u\), yielding the gated reward \(\tilde{r}_{i,g} = R_{\text{fmt}}(o_{i,g}) \cdot r_{i,g}\). The advantage is then normalized strictly within the group of rollouts generated for that specific query:

\[A_{i,g} = \frac{\tilde{r}_{i,g} - \mu_i}{\sigma_i + \epsilon_{\text{adv}}}, \quad \mu_i = \frac{1}{G} \sum_{g=1}^G \tilde{r}_{i,g}, \quad \sigma_i = \sqrt{\frac{1}{G} \sum_{g=1}^G (\tilde{r}_{i,g} - \mu_i)^2}\]

This group-relative formulation eliminates the need for an external value network or running global baselines, ensuring that policy gradients reflect the relative superiority of adversarial tactics for that particular query context.

4. Full-Sequence Token-Level GRPO Optimization: preserving the adversarial causal reasoning chain Reasoning-oriented VLMs generate internal thinking tokens prior to emitting visual and textual attack commands. Truncating the log-probability evaluation to only the final structured outputs \((t_{\text{img}}, u)\) breaks the causal chain of policy gradients and introduces mathematical bias into the importance sampling ratio \(\rho_{i,g,t}(\phi)\). MAMJ optimizes over the full sequence \(o = (\text{reasoning}, t_{\text{img}}, u)\), enforcing a token-level clipped surrogate objective with an unbiased \(k_3\) KL penalty against the fixed reference policy \(\pi_{\phi_0}\):

\[J_{\text{GRPO}}(\phi) = \frac{1}{|B|} \sum_{q_i \in B} \frac{1}{G} \sum_{g=1}^G \frac{1}{|o_{i,g}|} \sum_{t=1}^{|o_{i,g}|} \left[ \min\left( \rho_{i,g,t} A_{i,g}, \, \text{clip}(\rho_{i,g,t}, 1-\varepsilon_{\text{clip}}, 1+\varepsilon_{\text{clip}}) A_{i,g} \right) - \beta_{\text{KL}} D_{\text{KL}}^{i,g,t} \right]\]

Executed across 48 NVIDIA H100 GPUs with decoupled Megatron training, vLLM rollout generation, and Megatron-infer reference computation, this objective guarantees stable policy specialization without optimization collapse.

Loss & Training

Stage 2 training utilizes a balanced training set \(D_{\text{train}}\) of 1,000 queries spanning 13 safety categories. Optimization runs with AdamW at a peak learning rate of \(\eta = 1 \times 10^{-6}\) with cosine decay, weight decay \(10^{-2}\), and gradient clipping \(g_{\text{max}} = 1.0\). Batch parameters are set to \(|B|=32\) queries, \(G=4\) rollouts, inner epochs \(E=2\), clip range \(\varepsilon_{\text{clip}}=0.2\), and KL coefficient \(\beta_{\text{KL}}=0.01\). The vision encoder \(\phi_{\text{ViT}}\) is frozen, updating only the language stack \(\phi_{\text{train}}\) across 60 update steps (~12 hours total compute), achieving steady reward convergence and smooth policy entropy decay.

Key Experimental Results

Main Results

On the 13 safety scenarios of MM-SafetyBench (1,680 total evaluation samples), MAMJ was benchmarked against four template-based attacks and two state-of-the-art iterative attacks using an identical attacker base (Qwen3-VL-32B-Thinking) and image generator (Z-Image). Under a tighter query budget of only \(N_{\text{iter}}=2\), MAMJ substantially outpaces all baselines:

Target Victim VLM Attack Method Iteration Budget \(N_{\text{iter}}\) Average ASR (%) Gain over Baseline (pp)
GPT-4o (OpenAI) FigStep 1 14.40 -
GPT-4o (OpenAI) VisCo 4 54.23 -
GPT-4o (OpenAI) IDEATOR 3 61.85 Baseline
GPT-4o (OpenAI) MAMJ (Ours) 2 82.02 +20.17
Gemini-3-Pro-Preview (Google) FigStep 1 14.40 -
Gemini-3-Pro-Preview (Google) VisCo 4 47.98 -
Gemini-3-Pro-Preview (Google) IDEATOR 3 56.67 Baseline
Gemini-3-Pro-Preview (Google) MAMJ (Ours) 2 79.29 +22.62
Seed 2.0 (ByteDance) FigStep 1 22.50 -
Seed 2.0 (ByteDance) VisCo 4 55.77 -
Seed 2.0 (ByteDance) IDEATOR 3 62.92 Baseline
Seed 2.0 (ByteDance) MAMJ (Ours) 2 83.04 +20.12

Evaluating against four representative defense mechanisms on GPT-4o demonstrates that the learned artifact \((\theta^\star, \phi^\star)\) transfers zero-shot without any defense-specific retraining:

Defense Type & System VisCo (\(N_{\text{iter}}=4\)) IDEATOR (\(N_{\text{iter}}=3\)) MAMJ (\(N_{\text{iter}}=2\)) Improvement (pp)
Prompt Shield: AdaShield 39.40% 47.98% 52.68% +4.70
Prompt Shield: VLMGuard-R1 34.70% 42.74% 49.17% +6.43
Guardrail: Llama-Guard-4 39.58% 41.13% 54.17% +13.04
Guardrail: LlavaGuard 44.23% 49.64% 64.17% +14.53

Ablation Study

Extensive ablation experiments on GPT-4o systematically validate the necessity of each architectural component, critique mechanism, and normalization choice:

Ablation Dimension Configuration / Variant Average ASR (%) Empirical Takeaway
Optimization Axes Base attacker (unoptimized \(\theta_0, \phi_0\)) 35.4 ± 1.5 Unoptimized baseline performance
Optimization Axes ASP only (evolved \(\theta^\star\), frozen \(\phi_0\)) 68.7 ± 1.2 Strategy prompt contributes the largest single gain (+33.3 pp)
Optimization Axes RL only (GRPO from base prompt \(\theta_0\)) 63.2 ± 1.4 RL alone struggles under uninformative generic prompts
Optimization Axes Full MAMJ (two-axis staged optimization) 82.0 ± 0.9 Staged synergy achieves the highest success (+46.6 pp)
Diagnostic Feedback Sample-wise edits (no group aggregation) 56.4 ± 1.8 High variance from single-failure noise
Diagnostic Feedback Aggregator only (no per-trajectory Analyst) 62.8 ± 1.6 Lack of causal grounding weakens synthesized edits
Diagnostic Feedback MAMJ (Analyst attribution + Aggregator synthesis) 68.7 ± 1.2 Grounded group aggregation produces superior prompts
Reward Normalization Global running baseline 72.4 ± 1.5 Unstable due to reward scale drift across training
Reward Normalization Rollout-wise pooling (across all queries in batch) 76.8 ± 1.1 Mixing distinct queries dilutes query-specific signals
Reward Normalization Group-relative normalization (ours) 82.0 ± 0.9 Clean relative advantage within identical query contexts
Reasoning Log-Prob Truncated scope (outputs \((t_{\text{img}}, u)\) only) 73.5 ± 2.4 Policy drift and severe training instability around step 40
Reasoning Log-Prob Full sequence (reasoning + outputs) 82.02 ± 0.9 Preserves causal policy alignment and guarantees convergence

Key Findings

  • Meta-adaptation dominates per-sample iteration: Even when constrained to a strictly smaller per-query budget (\(N_{\text{iter}}=2\) vs. 3 or 4), MAMJ consistently outperforms iterative baselines across nearly all scenarios, especially in high-refusal categories like Legal Opinion (LO) and Hate Speech (HC).
  • Strong zero-shot parametric transferability: Although trained solely against GPT-4o, \((\theta^\star, \phi^\star)\) transfers seamlessly to unseen commercial victims (Gemini-3-Pro-Preview at 79.29% and Seed 2.0 at 83.04%) and cross-benchmark tasks (SafeBench at 87.80% ASR, +8.8 pp over IDEATOR).
  • Amortized computational cost: Despite an upfront offline training investment of \(6.36 \times 10^9\) equivalent LLM tokens (96.9% dominated by diffusion image synthesis), MAMJ reduces per-instance inference token costs by 33.4% compared to IDEATOR. The total cumulative compute crosses below IDEATOR's online search cost after 8,920 test queries.

Highlights & Insights

  • Lifting jailbreaking from sample content to the attacker: MAMJ reframes multimodal vulnerability discovery as meta-learning the attacker's policy, proving that static patterns are not the primary risk in multimodal safety.
  • Two-level decoupled critique: Decoupling per-trajectory evidence-based attribution from cross-trajectory strategy synthesis effectively filters out prompt optimization noise while preventing mode collapse.
  • CoT reasoning alignment in RL: Demonstrating that maintaining the reasoning chain within policy gradient computations is critical for multi-step adversarial planning models.

Limitations & Future Work

  • External Image Generator: The diffusion generator \(G\) (Z-Image) is treated as a frozen black box. While victim refusals stem primarily from textual/conceptual alignment rather than image rendering failures, end-to-end joint optimization of the visual synthesis backbone remains an open problem.
  • Evaluation Coverage: Benchmarks are primarily focused on MM-SafetyBench and SafeBench. Evaluating complex multi-image contexts, sequential visual dialogue, and embedded steganography represents important future red-teaming scope.
  • Defensive Implications: Standard guardrails and prompt shields suffer substantial degradation against meta-adaptive adversaries; defenses must transition toward meta-level adaptive monitors capable of identifying evolving adversarial strategies.
  • vs. Template-based attacks (FigStep, SIVA, HADES): Static methods freeze visual and textual patterns, making them fragile against layout and OCR filters; MAMJ evolves an adaptive, context-aware policy that bypasses explicit filters via symbolic metaphors.
  • vs. Sample-level iterative attacks (IDEATOR, VisCo): Iterative methods rewrite prompts per sample but keep iteration prompts and attacker weights static; MAMJ optimizes the strategy prompt and model parameters across trajectories, delivering higher success with fewer runtime turns.
  • vs. Text-only meta-jailbreaks (AutoDAN-Turbo, AdvPrompter): Pure-text methods operate in unimodal token spaces; MAMJ addresses the unique multimodal challenge of cross-modal synchronization between text payloads and diffusion generation.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ First framework to introduce two-axis meta-adaptive optimization to multimodal jailbreaking.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Evaluated across three leading commercial VLMs, four defenses, two benchmarks, with exhaustive ablations and cost analyses.
  • Writing Quality: ⭐⭐⭐⭐⭐ Rigorous methodology, clear mathematical derivations, and transparent disclosure of training dynamics.
  • Value: ⭐⭐⭐⭐⭐ Exposes a critical systemic vulnerability in frontier multimodal models, providing an indispensable foundation for next-generation defense design.