AgentTypo: Adaptive Typographic Prompt Injection Attacks against Black-box Multimodal Agents

arXiv:2510.04257 · cs.CR, cs.AI · Submitted 2025-10-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "AgentTypo: Adaptive Typographic Prompt Injection Attacks against Black-box Multimodal Agents".

Elias: Here's the gist of the paper. Multimodal agents built on large vision–language models (LVLMs) are increasingly deployed in open-world settings but remain highly vulnerable to prompt injection, especially through visual inputs <ref:2510.04257#pg2> AgentTypo, a black-box redteaming framework that mounts adaptive typographic prompt injection by embedding optimized text into webpage images <ref:2510.04257#pg7>.

Nadia: Framework Introduction The paper introduces AgentTypo, a black-box redteaming framework designed to expose typographic attack vulnerabilities in generalist LVLM agents in real-world scenarios under practical black-box settings <ref:2510.04257#pg4>.

Elias: This framework operates without access to model gradients or logits, making it deployable against black-box commercial agents <ref:2510.04257#pg7>.

Nadia: It utilizes a black-box Bayesian optimization algorithm to jointly refine the prompt position and properties, maximizing attack effectiveness while ensuring stealth from human detection <ref:2510.04257#pg7>.

Title and authors: Nadia: So this paper is talking about these large vision language models, the LVLMs. It says that even though these agents are being put into open world settings everywhere now, they are still really weak to prompt injection when we use visual inputs specifically.

Elias: Exactly. And what they introduce is a black-box redteaming framework called AgentTypo that mounts this adaptive typographic prompt injection by embedding optimized text directly into webpage images. It’s designed for scenarios where you can't actually see inside the model or get its internal workings, which makes it practical for real world applications.

Priya: So the core idea here seems to be finding ways to trick these agents visually, not just through text commands in a chat box. The framework operates without needing access to the model's gradients or logits, which is important because we can't just probe the model directly with gradient attacks.

Nadia: Right. And they use a black-box Bayesian optimization algorithm to jointly refine where that prompt goes and what its properties are—like font size, color, and contrast—to maximize how effective the attack is while trying to keep it hidden from human eyes.

Elias: That's the Automatic Typographic Prompt Injection algorithm, or ATPI. It tries to inject prompts into images by embedding them right into the image itself in a way that helps the model reconstruct those prompts successfully while also keeping it visually stealthy.

Priya: I noticed they use two losses to guide this process: one loss encourages successful prompt reconstruction by substituting captioners, and another loss measures how different the altered image looks compared to the original using LPIPS distance. That's a good way to balance getting the attack working and keeping it subtle.

Nadia: Then they build on this with AgentTypo-pro, which adds this strategy learning method inspired by continual learning. This means it iteratively improves the injection prompts based on evaluation feedback and retrieves successful past examples for what they call a strategy library.

Elias: It sounds like a closed loop where an Attacker LLM generates candidate hijacking prompts based on the attack objective, chat history, and strategies drawn from this library, while a Scoring LLM gives feedback to refine things further.

Priya: So the goal of this AgentTypo-pro system is to incrementally improve the hijacking effectiveness by reusing successful examples and extracted strategies in that continuous learning loop. That’s an interesting way to handle prompt evolution over time.

Nadia: The results they showed on the VWA-Adv benchmark are pretty strong. They found that for GPT-4o agents, this image-only attack can raise the success rate from twenty-three percent up to forty-five percent.

Elias: And that consistency across different models is what's compelling. They also showed in image plus text settings, AgentTypo achieves a sixty eight percent attack success rate on the GPT-4o model specifically <ref:2510.04257#pg2>.

Priya: Sixty eight percent is quite high, but we have to remember this was done in a controlled benchmark environment consisting of advertising, shopping, and Reddit domains. We need to be careful about generalizing those numbers to totally open world web environments where the context is much messier than that.

Nadia: Right. And they also showed that the combined integration of all three components—the ATPI algorithm, the strategy library, and the RAG module—produces what they claim is the strongest attack performance.

Elias: It really shows how layering these different techniques, from simple injection to adaptive optimization and then strategy learning, compounds the overall effectiveness against these multimodal systems.

Priya: A simple defense they mentioned is using a smaller captioning model to detect prompts inside images, which they found reduces the attack success rate on GPT-4o from that sixty eight percent down to twenty one percent <ref:2510.04257#pg2>. That's a measurable reduction in risk for any deployment right now.

Nadia: That trade-off the paper points out is that there’s an intrinsic tension between how stealthy you want the prompt to be and how effective it actually is; high success rates currently require the embedded text to be relatively conspicuous.

Elias: They also flagged a limitation on their evaluation setup, stating that they primarily conducted these experiments on three specific websites, and future work will need to extend those evaluations to a much broader range of web domains.

Priya: That’s a fair point about the generalization issue. So while this paper shows how potent this attack can be against GPT-4o agents, it also highlights that defenses might require more sophisticated detection methods than just a single captioning model looking at the image <ref:2510.04257#pg2>.

Nadia: This AgentTypo paper really puts the pressure on us to think about visual input security for multimodal agents in open world settings because they are clearly still very vulnerable.

Elias: It’s a good reminder that the attack surface isn't just what you type into a chat box anymore; it extends to how information is visually presented on a webpage.

Priya: I think the real implication here is that we need to develop more robust defenses that can handle these adaptive, visually optimized injections, rather than just looking for simple text patterns.

Nadia: So in summary, the AgentTypo paper successfully shows how to exploit visual weaknesses in LVLMs to compromise multiple commercial agents and it significantly improves the success rate of attacks against these systems.

Elias: It really demonstrates that typographic attacks pose a very real threat to LVLM agent systems right now, especially as people start trusting these AI agents more and more.

The paper's summary: Nadia: So, to recap, this paper introduces AgentTypo as a framework that lets researchers try to attack large vision language model agents in real world scenarios without ever being able to see inside the agent's code or its internal workings.

Elias: Right. It’s all about mounting adaptive typographic prompt injection by embedding optimized text into images, and the authors use a black-box Bayesian optimization algorithm to figure out where and how to put that text—things like font size or color—to get the best results while trying not to look suspicious.

Priya: From my side, I’m looking at the core mechanism called Automatic Typographic Prompt Injection, or ATPI. They use two losses to guide this process: one loss helps them figure out how well they can rebuild the prompt after it’s been injected by captioners, and another measures how much the altered image looks different from the original using a metric called LPIPS distance.

Nadia: That balancing act between getting a high success rate on that reconstruction loss and keeping the visual difference low with that stealthiness loss is what makes their attack strategy work in practice.

Elias: And then they take it further with AgentTypo-pro, which adds a strategy learning component inspired by continual learning. They create an Attacker LLM, a Scoring LLM, and a Summarizer LLM to build this closed-loop system where the prompt gets incrementally better based on feedback from successful examples.

Priya: So what this means for us is that we aren't just looking at static attacks anymore; we’re looking at systems that can learn how to evolve their visual defenses in response to continuous probing and exploitation.

Nadia: Exactly. The experimental results are pretty telling, especially on models like GPT-4o where the image-only attack success rate jumps up from twenty-three percent to forty-five percent <ref:2510.04257#pg2>.

Elias: And they show that when you combine all three parts—the ATPI algorithm, the strategy library, and that RAG module for continuous learning—you get what they claim is the strongest performance across various benchmarks.

Priya: The defense they proposed is pretty simple: use a smaller captioning model to look inside images for prompts, which manages to drop that success rate down from sixty eight percent to twenty-one percent on GPT-4o <ref:2510.04257#pg2>.

Nadia: That’s a significant drop in risk, but the authors also point out a limitation: there’s an inherent trade-off between how stealthy you want the attack to be and how effective it actually is right now; high success rates demand that the text embedded in the image looks pretty obvious.

Elias: And they also admit their evaluation was mostly on three specific websites, so extending those results to a much wider range of web domains is something they plan to do later.

Priya: It’s a real-world problem because users are increasingly trusting these AI agent systems, and this paper demonstrates that typographic attacks are a tangible threat to them in open world settings.

Nadia: So the bigger picture here is that we need much more robust defenses than just looking for simple text patterns when dealing with multimodal agents that process visual input.

The paper's improvements: Tom: So, we’re looking at how this paper suggests they can actually improve AgentTypo beyond just the basic attack setup.

Nadia: They introduce AgentTypo-pro, which incorporates a strategy learning method inspired by continual learning to make the prompt injection process smarter over time.

Elias: That means the system can iteratively improve the prompt based on injection strategies that get automatically identified by a summarizer LLM. It's like it learns from every successful attempt and uses that knowledge again in a closed-loop way.

Priya: From a privacy perspective, this strategy learning approach is interesting because it suggests an agent can adapt its attack parameters to maintain effectiveness against evolving defenses or models.

Nadia: Yeah, the whole pipeline now has an Attacker LLM generating prompts based on things like chat history and successful examples retrieved from a long-term memory called a strategy library.

Elias: So the Attacker LLM is constantly drawing from that library, and the Scoring LLM feeds back into it to adjust those strategies, which keeps the hijacking effectiveness increasing continuously.

Priya: It moves away from just finding one good configuration and toward a system that can maintain a high success rate even if the target model or environment changes slightly.

Nadia: And this whole structure—Attacker LLM, Scoring LLM, Summarizer LLM, and the Retrieval Augmented Generation module—is designed to create this continuous learning system for prompt hijacking.

Elias: It really shows how you can move from a one-time exploit attempt to something that tries to keep hitting targets successfully by constantly refining its approach.

Priya: The implication here is that for defense, we might have to look at detecting not just the current image, but patterns in how an agent is *learning* and adapting its visual responses.

Nadia: Right. This moves the focus from a static detection problem to a dynamic one, which makes it much harder for simple filters to catch.

Elias: It sets up some interesting questions about what kind of adversarial data you need to feed into that strategy library for the learning process to work effectively.

Conclusion: Tom: So we're wrapping up our look at AgentTypo: Adaptive Typographic Prompt Injection Attacks against Black-box Multimodal Agents.

Nadia: Basically, this paper shows how to exploit the visual weaknesses of large vision language model agents by embedding optimized text into images without needing any internal access to the model itself.

Elias: They show that using black-box Bayesian optimization allows for dynamically adjusting the prompt’s position and properties, like font size or color, to maximize the attack's effectiveness while trying to keep it visually subtle.

Priya: From what I’ve seen in their results on GPT-4o, that adaptive approach can raise the success rate from twenty-three percent up to forty-five percent in image only attacks <ref:2510.04257#pg2>.

Nadia: And they also show that when you put the whole thing together—the ATPI algorithm, the strategy library, and the RAG module—you get a sixty eight percent attack success rate in image plus text settings.

Elias: The real takeaway here is that typographic attacks are a very real threat to these multimodal agent systems right now as more people start trusting them.

Priya: I just want to say that the limitation they mention is important: there's an intrinsic trade-off where high attack success rates currently require the embedded text to be relatively conspicuous, which is something we have to consider for any real defense strategy.

Nadia: Right. So while they’ve shown how potent this attack can be against systems like GPT-4V and Gemini one point five Pro, it also points toward the urgent need for robust defenses that look beyond just simple text pattern detection when dealing with visual input security <ref:2510.04257#pg2>.

Elias: It really highlights that the attack surface isn't just what you type into a chat box anymore; it extends to how information is visually presented on a webpage.

Priya: I think the paper underscores that we need to develop more sophisticated detection methods than just looking for simple visual cues, especially as we look at those continuous learning improvements they introduced.

Nadia: Exactly. This AgentTypo paper proves that these models are still vulnerable, and it gives us a concrete framework for how to test those vulnerabilities in the real world without needing access to the model's internals.

Elias: It’s a solid piece of research because it demonstrates how layered attacks can compound effectiveness against systems like those based on vision language models.

Priya: I think this work paves the way for better privacy and measurement tools that can actually detect these visual manipulations before they cause real harm or data leakage.

Nadia: So that’s our look at AgentTypo: Adaptive Typographic Prompt Injection Attacks against Black-box Multimodal Agents. Thanks for tuning in.

Elias: We'll be right back with another paper coming up next time.

Yanjie Li, Yiming Cao, Dong Wang, Bin Xiao

Hong Kong Polytechnic University

cs.CR, cs.AI

Submitted: 2025-10-05

Updated: 2026-11-02

Comments: Submitted to IEEE TRANSACTIONS ON DEPENDABLE AND SECURE COMPUTING

Code: https://github.com/ddupont808/GPT-4V-Act

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 92/100

The gist: Multimodal agents built on large vision–language models (LVLMs) are increasingly deployed in open-world settings but remain highly vulnerable to prompt injection…

Key concepts

AgentTypo
A black-box redteaming framework used to test multimodal agents without access to their internal workings like model gradients. It automatically embeds optimized text into webpage images and uses Bayesian optimization to find the best way (position, font size, color) to hide the prompt while maximizing the chance of injection success.
ATPI Algorithm
The core attack mechanism that automatically injects prompts into images. It balances two goals: successfully fooling the vision-language model (prompt rebuilt loss) and keeping the resulting image looking visually similar to the original (stealthiness loss). This ensures the attack is both effective and hard to spot.
TPE Algorithm
A black-box Bayesian optimization method used by AgentTypo to find optimal parameters for the attack. It uses two probability distributions to intelligently search for the best combination of prompt characteristics, such as text position and contrast, maximizing the overall effectiveness of the injection strategy.

Terminology

Summary

Multimodal agents built on large vision–language models (LVLMs) are increasingly deployed in open-world settings but remain highly vulnerable to prompt injection, especially through visual inputs <ref:2510. p. 0

Where the paper says it

1 AgentTypo: Adaptive Typographic Prompt Injection Attacks against Black-box Multimodal Agents Yanjie Li, Yiming Cao, Dong Wang, Bin Xiao, Fellow, IEEE Abstract—Multimodal agents built on large vision–language models (LVLMs) are increasingly deployed in open-world settings but remain highly vulnerable to prompt injection, especially through visual inputs. We introduce AgentTypo, a black-box redteaming framework that mounts adaptive typographic prompt injection by embedding optimized text into webpage images.

04257#pg2> AgentTypo, a black-box redteaming framework that mounts adaptive typographic prompt injection by embedding optimized text into webpage images <ref:2510. p. 0
Where the paper says it

1 AgentTypo: Adaptive Typographic Prompt Injection Attacks against Black-box Multimodal Agents Yanjie Li, Yiming Cao, Dong Wang, Bin Xiao, Fellow, IEEE Abstract—Multimodal agents built on large vision–language models (LVLMs) are increasingly deployed in open-world settings but remain highly vulnerable to prompt injection, especially through visual inputs. We introduce AgentTypo, a black-box redteaming framework that mounts adaptive typographic prompt injection by embedding optimized text into webpage images.

04257#pg7>.

Framework Introduction

The paper introduces AgentTypo, a black-box redteaming framework designed to expose typographic attack vulnerabilities in generalist LVLM agents in real-world scenarios under practical black-box settings <ref:2510.04257#pg4>. This framework operates without access to model gradients or logits, making it deployable against black-box commercial agents <ref:2510. p. 1

Where the paper says it

Our key contributions are summarized as follows: • We propose AgentTypo, the first red-teaming framework to evaluate and expose typographic attack vulnerabilities in generalist LVLM agents in real-world scenarios under practical black-box settings. AgentTypo operates without access to model gradients or logits, making the method deployable against black-box commercial agents.

04257#pg7>. It utilizes a black-box Bayesian optimization algorithm to jointly refine the prompt position and properties, maximizing attack effectiveness while ensuring stealth from human detection <ref:2510. p. 1
Where the paper says it

• To improve both attack success rate and stealthiness, we employ a black-box Bayesian optimization algorithm to jointly refine the prompt position and properties. This enhances attack effectiveness while ensuring stealth from human detection.

04257#pg7>.

Attack Mechanism

The Automatic Typographic Prompt Injection (ATPI) algorithm is the core mechanism of AgentTypo, aiming to inject prompts into images to maximize attack success rate while preserving transferability and visual stealthiness <ref:2510.04257#pg5>. The ATPI algorithm exploits latent vulnerabilities by embedding adversarial prompts directly into the image <ref:2510. p. 3

Where the paper says it

The overall pipeline of the ATPI algorithm is shown in Figure 3. The proposed ATPI algorithm exploits the latent vulnerabilities in LVLM agents by embedding adversarial prompts directly into the image.

04257#pg5>. It achieves this by jointly minimizing two losses: the prompt rebuilt loss, which encourages successful reconstructions of the injected prompts by substituting vision-language models <ref:2510. p. 4
Where the paper says it

, font size and color) to maximize the adversarial effect and minimize visual disruption and attack an ensemble of visual language models to improve the transferability. where Lprompt rebuilt encourages successful reconstructions of the injected prompts by substitute vision-language models, Lstealthiness ensures visual inconspicuousness, and λ is a hyperparameter that controls the trade-off between the two objectives.

04257#pg4>, and the stealthiness loss, which measures perceptual differences between original and altered images using LPIPS distance <ref:2510. p. 4
Where the paper says it

b) The Stealthiness Loss of ATPI: To improve the visual stealthiness of the inserted prompt, we apply the Learned Perceptual Image Patch Similarity (LPIPS) [58] distance to measure the distance between the original image and the altered image. LPIPS quantifies the perceptual differences between two images, as judged by deep neural networks trained to align with human visual perception.

04257#pg4>.

Adaptive Optimization

The black-box optimization problem is solved using the Tree-structured Parzen Estimator (TPE) algorithm, a variant of Bayesian Optimization methods <ref:2510.04257#pg7>. TPE employs two separate probability distributions to distinguish between good and bad hyperparameters, choosing the configuration x that maximizes the ratio l(x)/g(x) based on the acquisition function P(y ≤ yγx, D) <ref:2510. p. 5

Where the paper says it

The optimization objectiveness can be formulated as xopt ∈ argmin x∈X LATPI. (5) TPE employs two separate probability distributions to distinguish between good and bad hyperparameters.

04257#pg7>. This process iteratively adjusts parameters such as text position, font size, color, and contrast to maximize the objective LATPI = Lprompt rebuilt + λLstealthiness <ref:2510. p. 3
Where the paper says it

, font size, contrast, color, location) to maximize the adversarial effect and minimize visual disruption and attack an ensemble of visual language models to improve the transferability. a) The Prompt Rebuilt Loss of ATPI: Given an image target and an adversarial prompt P to be inserted, the aim of ATPI is to ensure that the adversarial prompt is effectively reconstructed by target vision-language models.

04257#pg4>.

Adaptive Prompt Optimization

AgentTypo-pro further enhances the attack by incorporating a novel adaptive approach based on strategy learning inspired by continual learning <ref:2510.04257#pg5>. This system consists of an Attacker LLM, a Scoring LLM, a Summarizer LLM, and a Retrieval-Augmented Generation (RAG) module for continuous learning <ref:2510. p. 5

Where the paper says it

18: Return xopt, I ∗ , and the evaluation score. The pipeline consists of four primary components: an Attacker LLM, a Scoring LLM, a Summarizer LLM, and a Retrieval-Augmented Generation (RAG) module for continuous learning.

04257#pg5>. The Attacker LLM generates candidate hijacking prompts based on the attack objective ϕ, chat history Ht, successful examples retrieved from long-term memory, and a sampled strategy ψt drawn from the strategy library Ψ <ref:2510. p. 6
Where the paper says it

As illustrated in Figure 4, at t-th iteration, its input consists of several components: the attack objective ϕ, expressed as a natural language description of the desired action; the chat history Ht from previous steps, where each step contains the inserted prompt, the agent’s output action, and the corresponding evaluation score; the top-k successful examples {e1, . , ek} retrieved from long-term memory; and a sampled strategy ψt drawn from the strategy library Ψ.

04257#pg5>.

Experimental Results

Experiments on the VWA-Adv benchmark show that AgentTypo significantly outperforms most recent text-based and image-based prompt injection attacks <ref:2510.04257#pg7>. For GPT-4o agents, the image-only attack raises the success rate from 23% to 45%, with consistent results across GPT-4V, GPT-4o-mini, Gemini 1. p. 0

Where the paper says it

Experiments on the VWA-Adv benchmark across Classifieds, Shopping, and Reddit scenarios show that AgentTypo significantly outperforms the latest image-based attacks such as AgentAttack. On GPT-4o agents, our image-only attack raises the success rate from 23% to 45%, with consistent results across GPT-4V, GPT-4o-mini, Gemini 1.

5 Pro, and Claude 3 Opus <ref:2510. p. 0
Where the paper says it

On GPT-4o agents, our image-only attack raises the success rate from 23% to 45%, with consistent results across GPT-4V, GPT-4o-mini, Gemini 1. 5 Pro, and Claude 3 Opus.

04257#pg7>. In image+text settings, AgentTypo achieves 68% ASR on the GPT-4o model <ref:2510. p. 1
Where the paper says it

Moreover, we evaluate our method using the real-world agent attack benchmark VWA-Adv [15], which includes three domains: advertising, shopping, and social networking, and 77 distinct adversarial tasks. Our AgentTypo-pro attack achieves the highest success rates compared to the state-of-the-art image-based prompt injection attack, AgentAttack [15], increasing the attack success rate (ASR) from 23% to 45% for image-only attacks and from 26% to 68% for text+image attacks on the GPT-4o model.

04257#pg7>. The combined integration of all three components (ATPI, strategy library, and RAG) produces the strongest attack performance <ref:2510. p. 9
Where the paper says it

, it improves 12% on GPT-4o-mini compared with Strategy-only method). The combined integration of all three components (ATPI, strategy library, and RAG) produces the strongest attack performance, demonstrating the strong complementarity between visual vulnerability exploitation and prompt optimization.

04257#pg7>.

Defense and Limitations

A simple defense proposed is using a smaller captioning model to detect prompts within images, which reduces the attack success rate on GPT-4o from 0.68 to 0.21 <ref:2510. p. 0

Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

04257#pg8>. However, there is an intrinsic trade-off between stealth and effectiveness where high attack success rates currently require the embedded text to be relatively conspicuous <ref:2510. p. 0
Where the paper says it

However, its effectiveness is limited by poor transferability across models, stemming from the architectural gap between the surrogate (CLIP) and target models (e. Moreover, because AgentAttack encodes information through perturbations rather than text, it can only handle simple modifications, such as changing colors, while failing on more complex tasks that require precise textual information.

04257#pg9>. Future work will explore methods to improve inconspicuity by using advanced image generation techniques <ref:2510.04257#pg9>. The empirical evaluation was conducted primarily on three websites, and future work will extend these evaluations to a broader range of web domains <ref:2510.04257#pg9>.

Conclusion

In summary, AgentTypo successfully exploits LVLM’s visual weaknesses to compromise multiple commercial multimodal agents and significantly improves the success rate of attacks against these systems <ref:2510.04257#pg10>. The findings highlight the urgent need for robust defenses, especially as users increasingly trust AI agent systems <ref:2510. p. 1

Where the paper says it

Our AgentTypo-pro attack achieves the highest success rates compared to the state-of-the-art image-based prompt injection attack, AgentAttack [15], increasing the attack success rate (ASR) from 23% to 45% for image-only attacks and from 26% to 68% for text+image attacks on the GPT-4o model. To the best of our knowledge, AgentTypo is the first framework to expose typographic attack vulnerabilities in LVLM agents, highlighting urgent needs for the development of robust defenses in multimodal agent systems.

04257#pg10>. The paper demonstrates that typographic attacks pose a real-world threat to LVLM agent systems <ref:2510. p. 1
Where the paper says it

To the best of our knowledge, AgentTypo is the first framework to expose typographic attack vulnerabilities in LVLM agents, highlighting urgent needs for the development of robust defenses in multimodal agent systems. Our key contributions are summarized as follows: • We propose AgentTypo, the first red-teaming framework to evaluate and expose typographic attack vulnerabilities in generalist LVLM agents in real-world scenarios under practical black-box settings.

04257#pg10>.

References

[1] S. Wu, H. Fei, L. Qu, W. Ji, and T.-S. Chua, “Next-gpt: Any-to-any multimodal llm,” in Forty-first International Conference on Machine Learning, 2024 <ref:2510. p. 7

Where the paper says it

• AdvAgent [21]: AdvAgent employs Supervised Finetuning (SFT) and reinforcement learning to train a lightweight adversarial prompting model based on feedback from a black-box agent. In the training phase, we first generate some success examples using GPT-4o, and then use SFT and DPO to finetune the Mistral-7BInstruct-v0.

04257#pg10>.

[6] B. Zheng, B. Gou, J. Kil, H. Sun, and Y. Su, “Gpt-4v(ision) is a generalist web agent, if grounded,” in Proceedings of the 41st International Conference on Machine Learning, ser. p. 7

Where the paper says it

• AdvAgent [21]: AdvAgent employs Supervised Finetuning (SFT) and reinforcement learning to train a lightweight adversarial prompting model based on feedback from a black-box agent. In the training phase, we first generate some success examples using GPT-4o, and then use SFT and DPO to finetune the Mistral-7BInstruct-v0.

ICML’24 <ref:2510. p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

04257#pg10>.

[15] C. H. Wu, R. R. Shah, J. Y. Koh, R. Salakhutdinov, D. Fried, and A. Raghunathan, “Dissecting adversarial robustness of multimodal LM agents,” in The Thirteenth International Conference on Learning Representations <ref:2510. p. 0

Where the paper says it

Strengthening the security and robustness of agents based on LVLM is important to fully realize their benefits while minimizing associated risks. To examine traditional and multimodal agent weaknesses, recent research has explored both text-based and image-based adversarial attacks [15], [17]–[22].

04257#pg10>.

[61] A. Mehrotra, M. Zampetakis, P. Kassianik, B. Nelson, H. Anderson, Y. Singer, and A. Karbasi, “Tree of attacks: Jailbreaking black-box llms automatically,” Advances in Neural Information Processing Systems <ref:2510.04257#pg10>.

[64] OpenAI, “Gpt-4 technical report,” arXiv preprint arXiv: 2303. p. 0

Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

08774, 2024 <ref:2510. p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

04257#pg10>.

[68] A. Q. Jiang, A. Sablayrolles, and A. Mensch, “Mistral 7b,” 2023 <ref:2510. p. 0

Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308. p. 0

Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

14132, 2023 <ref:2510. p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

04257#pg10>.

[59] S. Watanabe, “Tree-structured parzen estimator: Understanding its algorithm components and their roles for better empirical performance,” 2023 <ref:2510. p. 0

Where the paper says it

We introduce AgentTypo, a black-box redteaming framework that mounts adaptive typographic prompt injection by embedding optimized text into webpage images. Our automatic typographic prompt injection (ATPI) algorithm maximizes prompt reconstruction by substituting captioners while minimizing human detectability via a stealth loss, with a Tree-structured Parzen Estimator guiding black-box optimization over text placement, size, and color.

04257#pg10>.

[67] Q. Zhan, Z. Liang, Z. Ying, and D. Kang, “Injecagent: Benchmarking indirect prompt injections in tool-integrated large language model agents,” in Findings of the Association for Computational Linguistics ACL 2024 <ref:2510. p. 0

Where the paper says it

Our findings reveal that AgentTypo poses a practical and potent threat to multimodal agents and highlight the urgent need for effective defense. Index Terms—Large Vision Language Model, Agent, Prompt Injection Attack, LLM Security I.

04257#pg10>.

[68] A. Q. Jiang, A. Sablayrolles, and A. Mensch, “Mistral 7b,” 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[58] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in CVPR, 2018 <ref:2510. p. 4

Where the paper says it

b) The Stealthiness Loss of ATPI: To improve the visual stealthiness of the inserted prompt, we apply the Learned Perceptual Image Patch Similarity (LPIPS) [58] distance to measure the distance between the original image and the altered image. LPIPS quantifies the perceptual differences between two images, as judged by deep neural networks trained to align with human visual perception.

04257#pg10>.

[63] X. Liu, P. Li, G. E. Suh, Y. Vorobeychik, Z. Mao, S. Jha, P McDaniel, H Sun, B Li and C Xiao “AutoDAN-turbo: A lifelong agent for strategy self-exploration to jailbreak LLMs” in The Thirteenth International Conference on Learning Representations <ref:2510. p. 5

Where the paper says it

ADAPTIVE PROMPT OPTIMIZATION BASED ON CONTINUAL LEARNING To further enhance the success rate of prompt injection, we introduce a novel, adaptive approach for optimizing hijacking prompts, inspired by the concept of continual learning [60]. Adaptive methods have been used in jailbreaking LLMs, such as Tree of Attack [61], AutoDAN [62], and AutoDAN-turbo [63].

04257#pg10>.

[60] L. Wang, X. Zhang, H. Su, and J. Zhu, “A comprehensive survey of continual learning: Theory, method and application,” IEEE Transactions on Pattern Analysis and Machine Intelligence <ref:2510.04257#pg10>.

[47] K. Greshake, S. p. 2

Where the paper says it

Most existing works focus on text-based indirect prompt injection, in which adversarial prompts are embedded within web content, typically via HTML [20]–[22], [46]. Greshake et al.

Abdelnabi, S. Mishra, C. Endres, T. Holz, and M. Fritz “Not what you’ve signed up for: Compromising real-world llm-integrated applications with indirect prompt injection” 2023 <ref:2510. p. 2
Where the paper says it

Greshake et al. [47] propose the first indirect prompt injection attack, where the adversaries remotely exploit the LLM applications by injecting the prompts into data likely to be retrieved.

04257#pg10>.

[48] X. Shen, Z. Chen, M. Backes, Y. Shen, and Y. Zhang “do anything now”: Characterizing and evaluating in-the-wild jailbreak prompts on large language models” 2024 <ref:2510. p. 2

Where the paper says it

Second, our focus is on prompt injection within the agent, which targets different objectives and outcomes compared to traditional jailbreaking attacks. Typographic Attacks for Large Language Models Unlike gradient-based optimization attacks that add suffixes to prompts or introduce perturbations to images to result in malicious text outputs such as GCG [23] and PGD [52], [53], typographic attacks [54]–[56] manipulate textual elements within images, presenting a novel and significant threat to large vision language models (LVLMs).

04257#pg10>.

[54] H. Cheng, E. Xiao, J. Gu, L. Yang, J. Duan, J Zhang, J Cao K Xu and R Xu “Unveiling typographic deceptions: Insights of the typographic vulnerability in large vision-language models” in European Conference on Computer Vision Springer 2024 <ref:2510. p. 0

Where the paper says it

1 AgentTypo: Adaptive Typographic Prompt Injection Attacks against Black-box Multimodal Agents Yanjie Li, Yiming Cao, Dong Wang, Bin Xiao, Fellow, IEEE Abstract—Multimodal agents built on large vision–language models (LVLMs) are increasingly deployed in open-world settings but remain highly vulnerable to prompt injection, especially through visual inputs. We introduce AgentTypo, a black-box redteaming framework that mounts adaptive typographic prompt injection by embedding optimized text into webpage images.

04257#pg10>.

[56] Y. Cao, Y. Xing, J Zhang, D Lin, T Zhang, I Tsang, Y Liu and Q Guo “Scenetap: Scene-coherent typographic adversarial planner against vision-language models in real-world environments” in Proceedings of the Computer Vision and Pattern Recognition Conference 2025 <ref:2510. p. 2

Where the paper says it

90% on classification tasks through targeted interventions. Additionally, SceneTAP [56] introduces a novel framework for generating scene-coherent typographic adversarial attacks that effectively mislead LVLMs while preserving visual naturalness, showcasing practical applications in real-world environments.

04257#pg10>.

[38] S. Zhou, F. F. Xu, H Zhu, X Zhou, R Lo, A Sridhar X Cheng T Ou Y Bisk D Fried et al “Webarena: A realistic web environment for building autonomous agents” arXiv preprint arXiv:2307. p. 0

Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

13854 2023 <ref:2510. p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

04257#pg10>.

[46] F. Wu, S. Wu, Y. Cao, and C Xiao “Wipi: A new web threat for llm-driven web agents” 2024 <ref:2510. p. 2

Where the paper says it

As a result, ensuring the robustness and trustworthiness of multimodal agents remains an open and important research challenge. Prompt Injection Attacks in LLM Agents Although web agents significantly improve user convenience and operational efficiency, they also introduce new security threats [17], [43]–[45].

04257#pg10>.

[37] S. Yao, H Chen, J Yang, and K Narasimhan “Webshop: Towards scalable real-world web interaction with grounded language agents” Advances in Neural Information Processing Systems <ref:2510. p. 7

Where the paper says it

Experiment Setup a) Evaluation Benchmark: We conduct experiments on the VWA-Adv benchmark [15], a real-world multimodal adversarial attack evaluation benchmark built upon the normal tasks from VisualWebArena [40]. VisualWebArena is a benchmark specifically designed to evaluate the performance of multimodal agents on realistic, visually grounded web-based tasks, using annotated webpage screenshots and text as input.

04257#pg10>.

[40] J. Y. Koh, R Lo, L Jang V Duvvur M C Lim P-Y Huang G Neubig S Zhou R Salakhutdinov and D Fried “Visualwebarena: Evaluating multimodal agents on realistic visual web tasks” ACL 2024 <ref:2510. p. 7

Where the paper says it

Experiment Setup a) Evaluation Benchmark: We conduct experiments on the VWA-Adv benchmark [15], a real-world multimodal adversarial attack evaluation benchmark built upon the normal tasks from VisualWebArena [40]. VisualWebArena is a benchmark specifically designed to evaluate the performance of multimodal agents on realistic, visually grounded web-based tasks, using annotated webpage screenshots and text as input.

04257#pg10>.

[39] A. Patel, M. Hofmarcher, C Leoveanu-Condrei, M.-C. Dinu C Callison-Burch and S Hochreiter “Large language models can selfimprove at web agent tasks” arXiv preprint arXiv:2405. p. 2

Where the paper says it

Generalist Web Agents As large language models (LLMs) and large vision-language models (LVLMs) continue to advance, their capabilities in visual perception, language understanding [29], reasoning [30], [31], and following instructions [32], [33] are consistently improving. This progress has facilitated the widespread application of these models in developing intelligent agents for complex tasks, such as automated web navigation and interaction [34]–[36].

20309 2024 <ref:2510. p. 0
Where the paper says it

To address these limitations, we propose AgentTypo-base and AgentTypo-pro, two novel red-teaming approaches aimed at exposing multimodal vulnerabilities in LVLM agents. The AgentTypo-base attack utilizes vulnerabilities in the image modality and uses Bayesian optimization to optimize the insertion of prompts and other properties, maximizing attack effectiveness arXiv:2510.

04257#pg10>.

[57] D. Dupont, “GPT-4V-Act: Chromium Copilot,” https://github. p. 3

Where the paper says it

This agent paradigm leverages the reasoning and vision capabilities of LVLMs to complete sophisticated web tasks. In this paper, we adopt the agent setup consistent with VisualWebArena [40] and GPT-4V-Act [57].

com/ddupont808/GPT-4V-Act, 2024 <ref:2510. p. 3
Where the paper says it

This agent paradigm leverages the reasoning and vision capabilities of LVLMs to complete sophisticated web tasks. In this paper, we adopt the agent setup consistent with VisualWebArena [40] and GPT-4V-Act [57].

04257#pg10>.

[66] Anthropic, “The claude 3 model family: Opus, sonnet, haiku,” 2024 <ref:2510. p. 0

Where the paper says it

On GPT-4o agents, our image-only attack raises the success rate from 23% to 45%, with consistent results across GPT-4V, GPT-4o-mini, Gemini 1. 5 Pro, and Claude 3 Opus.

04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M. Kamfonas, “Detecting language model attacks with perplexity,” arXiv preprint arXiv:2308.14132, 2023 <ref:2510.04257#pg10>.

[69] G. Alon and M.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Adaptive Typographic Prompt Injection (AgentTypo) can be used to mount attacks by embedding optimized text into webpage images. This allows agents to be manipulated even in black-box settings without needing internal model access.

  2. The ATPI algorithm uses a black-box Bayes Optimization algorithm to "dynamically adjust the placement and characteristics of the prompt (e.g., font size, contrast, color) to maximize the adversarial effect and minimize visual disruption and attack an ensemble of visual language models to improve the transferability." This enables finding optimal text properties for stealth while maximizing success rates.

  3. AgentTypo-pro introduces a strategy learning method that iteratively improves the prompt based on injection strategies automatically identified by a summarizer LLM. This allows the system to incrementally improve hijacking effectiveness by reusing successful examples and extracted strategies in a closed-loop system.

  4. The overall AgentTypo-pro pipeline, incorporating an Attacker LLM, Scorer LLM, Retrieval-Augmented Generation (RAG), and Summarizer LLM, enables a closed-loop system where the Attacker LLM generates prompts based on strategies and feedback from a Scoring LLM.

  5. The system can be deployed to achieve high Attack Success Rates (ASR) on GPT-4o agents, reaching up to 68% in image+text settings across diverse benchmarks like Classifieds, Shopping, and Reddit.

Abstract

Multimodal agents built on large vision-language models are increasingly deployed in open-world settings but remain highly vulnerable to prompt injection. Typographic text-in-image attacks such as FigStep have been evaluated mainly on standalone vision-language models. For agents, user goals, webpage context, and action history make the effect of embedded instructions uncertain, and robustness to behavior-guided, adaptive visual injections remains insufficiently studied. We introduce AgentTypo, a black-box red-teaming framework that launches adaptive typographic prompt injection by embedding optimized text into webpage images. We formulate an attention-guided automatic typographic prompt-injection objective that favors the attacker's expected answer over the clean-page answer, together with task-conditioned visual attention from small question-answering models. A novel black-box optimization method searches text position, contrast, and color. To further strengthen the attack, we develop a multi-LLM system that iteratively refines injection prompts using evaluation feedback and retrieves successful past examples for strategy learning. Experiments on the VWA-Adv and Mind2Web benchmarks show that AgentTypo significantly outperforms state-of-the-art image-based prompt injection attacks. Across seven closed- and open-source model backends, AgentTypo-pro achieves 45%-83% ASR, outperforming the CLIP-based AgentAttack by 29-47% and previous typographic attacks by 25%-37% on VWA-Adv.

Sources

Related papers