Connected & Protected
24/7 Support
HomeBlogNew image-based prompt injection attack targets multimodal AI models
Artificial Intelligence, Cyberattacks, Generative AI, Security
#Artificial Intelligence, Cyberattacks, Generative AI, Security
#image-based prompt injection
#multimodal AI security
#CrossMPI technique
#AI model vulnerabilities
#generative AI risks
#vision-language models
#enterprise AI security
#cybersecurity threats

New image-based prompt injection attack targets multimodal AI models

Security researchers have developed a new image-based prompt injection attack that can manipulate how multimodal AI systems interpret user instructions without modifying the original text prompt, potentially expanding security risks for AI agents and vision-language systems. In a research paper published this week, researchers from Xidian University described a technique called “CrossMPI,” which uses nearly imperceptible image perturbations to alter how large vision-language models (LVLMs) process both visual and textual inputs. “CrossMPI can steer the model’s interpretation of both textual and visual inputs via image-only prompt injection,” the researchers wrote in the paper. Unlike traditional prompt injection attacks, which typically rely on malicious text instructions embedded in prompts or webpages, the new technique attempts to change how the model interprets a benign user request by manipulating images alone. “The perturbed image can manipulate the model’s understanding of the user’s instruction,” the paper said. In one example described in the paper, researchers subtly modified an image of an airplane using nearly imperceptible pixel-level perturbations invisible to human users. When a multimodal AI system was then asked whether the airplane belonged to Air Canada, the manipulated image caused the model to incorrectly identify the object as “a mobile phone,” illustrating how the attack could distort both visual understanding and interpretation of the user’s task. The findings add to growing concerns around multimodal AI security as enterprises increasingly deploy AI copilots, autonomous agents, document-processing assistants, and vision-enabled workflows that combine image and text reasoning. Apeksha Kaushik, senior principal analyst at Gartner, said the risks could grow rapidly as enterprises adopt more multimodal AI systems. “By 2030, 80% of enterprise software and applications will be multimodal, up from 1% in 2024,” Kaushik said. Attack targets multimodal reasoning layers Prompt injection has emerged as one of the most closely watched risks in generative AI systems, particularly as organizations adopt AI agents capable of interacting with enterprise applications, websites, documents, and external tools. Most existing prompt injection attacks rely on malicious text embedded in prompts, webpages, or hidden instructions. Some multimodal attacks have also attempted to manipulate AI behavior using images containing visible or hidden text instructions. The researchers argued their approach differs because it attempts to alter how the model interprets the original task itself through image perturbations alone. By contrast with earlier methods, the researchers noted that CrossMPI uses image modifications to “change the model’s interpretation of both the visual and textual prompts.” The paper said the attack specifically targets the “hidden state space of LVLMs” — the stage where models combine textual instructions and visual evidence into internal representations before generating outputs. According to the paper, the most effective attack layers were not the final output layers traditionally targeted in adversarial AI attacks, but intermediate layers where visual and textual information are fused together. Researchers claim strong black-box transferability The researchers evaluated the technique against multiple open-source LVLMs, including MiniGPT4, BLIP-2, InstructBLIP, BLIVA, and Qwen2.5-VL, the paper added. According to the paper, the attack achieved an average success rate of 66.36% across tested models, outperforming prior baseline attacks by roughly 41 percentage points. The researchers also said the technique demonstrated “strong transferability in black-box settings,” meaning the attacks remained effective even without direct access to a target model’s parameters or architecture. The paper further claimed the perturbations remained visually stealthy while maintaining effectiveness across multiple LVLM architectures. No effective defense The researchers evaluated several defense mechanisms designed to neutralize hidden image manipulations, including random resizing, image rotation, JPEG compression, and inference-level safeguards such as SmoothVLM, a specialized defense framework designed to protect Vision-Language Models (VLMs) from patched visual prompt injections, and DPS, which guides models using partial image views. According to the paper, SmoothVLM proved the most effective, reducing attack success rates to below 5% in several scenarios, while JPEG compression also weakened the attacks by suppressing high-frequency image artifacts. However, the researchers said none of the tested defenses completely eliminated the attacks, suggesting stronger multimodal AI security protections may still be needed. Enterprise AI deployments may widen exposure The research arrives as enterprises rapidly expand deployments of multimodal AI systems capable of processing screenshots, PDFs, dashboards, forms, video streams, and enterprise documents alongside natural language prompts. The researchers noted that adversarial examples generated using the technique could potentially “mislead VLM-based web agents” and “disrupt real-world object detectors.” “Even if textual inputs are sanitized, manipulated images can still subvert the model’s outputs or actions,” Kaushik said. She said organizations that use multimodal AI for document processing, customer interactions, content moderation, and autonomous systems may face increasing exposure to adversarial image manipulation and prompt injection attacks. “Security controls designed for unimodal systems are insufficient,” Kaushik said. The researchers acknowledged that the work was conducted in controlled research settings using open-source models and did not describe observed exploitation in real-world enterprise environments.

M
Microsoft disrupts malware code-signing service used by ransomware gangs | CSO Online
Content Author
Content Specialist

With extensive experience in cybersecurity and AI technologies, the author specializes in analyzing emerging threats to AI systems and providing insights for businesses on safeguarding their multimodal AI applications.

Expertise: Cybersecurity, Artificial Intelligence, Prompt Injection Attacks, Multimodal AI, Risk Management, Generative AI, Threat Analysis
Published: May 20, 2026
8 min read

Quick Answer

A new technique called 'CrossMPI' allows attackers to manipulate multimodal AI models by altering images, potentially distorting the AI's interpretation of user prompts without changing the text.

TL;DR

Researchers have unveiled a novel image-based prompt injection attack known as 'CrossMPI' that targets multimodal AI systems. This method enables attackers to influence how AI interprets both visual and textual inputs, raising significant security concerns.

Key Takeaways

  • 1CrossMPI uses imperceptible image perturbations to alter AI interpretations.
  • 2The attack manipulates user instructions via images without changing text prompts.
  • 3Growing adoption of multimodal AI systems increases vulnerability to such attacks.
New image-based prompt injection attack targets multimodal AI models
  1. Analytics
  2. Application Security
  3. Artificial Intelligence
  4. Business Continuity
  5. Business Operations
  6. Careers
  7. Cloud Security
  8. Compliance
  9. Critical Infrastructure
  10. Cybercrime
  11. Enterprise Buyer’s Guides
  12. Generative AI
  13. Identity and Access Management
  14. Industry
  15. IT Leadership
  16. IT Management
  17. Network Security
  18. Physical Security
  19. Privacy
  20. Risk Management
  21. Security
  22. Security Infrastructure
  23. Software Development
  24. Vulnerabilities

New image-based prompt injection attack targets multimodal AI models

Researchers say the technique can manipulate how vision-language models interpret both images and user prompts.

AI, Prompt, Woman, Engineer
Credit: Irina Strelnikova - shutterstock.com

Security researchers have developed a new image-based prompt injection attack that can manipulate how multimodal AI systems interpret user instructions without modifying the original text prompt, potentially expanding security risks for AI agents and vision-language system

CrowdStrike
Related Product

CrowdStrike Advanced Defend

by CrowdStrike$15.41/mo

CrowdStrike provides advanced cloud-native platform protection for endpoints, cloud workloads, identities, and data with a single lightweight agent.

Learn More
s.

In a research paper published this week, researchers from Xidian University described a technique called “CrossMPI,” which uses nearly imperceptible image perturbations to alter how large vision-language models (LVLMs) process both visual and textual inputs.

“CrossMPI can steer the model’s interpretation of both textual and visual inputs via image-only prompt injection,” the researchers wrote in the paper.

Unlike traditional prompt injection attacks, which typically rely on malicious text instructions embedded in prompts or webpages, the new technique attempts to change how the model interprets a benign user request by manipulating images alone.

“The perturbed image can manipulate the model’s understanding of the user’s instruction,” the paper said.

In one example described in the paper, researchers subtly modified an image of an airplane using nearly imperceptible pixel-level perturbations invisible to human users. When a multimodal AI system was then asked whether the airplane belonged to Air Canada, the manipulated image caused the model to incorrectly identify the object as “a mobile phone,” illustrating how the attack could distort both visual understanding and interpretation of the user’s task.

The findings add to growing concerns around multimodal AI security as enterprises increasingly deploy AI copilots, autonomous agents, document-processing assistants, and vision-enabled workflows that combine image and text reasoning.

Apeksha Kaushik, senior principal analyst at Gartner, said the risks could grow rapidly as enterprises adopt more multimodal AI systems.

“By 2030, 80% of enterprise software and applications will be multimodal, up from 1% in 2024,” Kaushik said.

Attack targets multimodal reasoning layers

Prompt injection has emerged as one of the most closely watched risks in generative AI systems, particularly as organizations adopt AI agents capable of interacting with enterprise applications, websites, documents, and external tools.

Most existing prompt injection attacks rely on malicious text embedded in prompts, webpages, or hidden instructions. Some multimodal attacks have also attempted to manipulate AI behavior using images containing visible or hidden text instructions.

The researchers argued their approach differs because it attempts to alter how the model interprets the original task itself through image perturbations alone.

By contrast with earlier methods, the researchers noted that CrossMPI uses image modifications to “change the model’s interpretation of both the visual and textual prompts.”

The paper said the attack specifically targets the “hidden state space of LVLMs” — the stage where models combine textual instructions and visual evidence into internal representations before generating outputs.

According to the paper, the most effective attack layers were not the final output layers traditionally targeted in adversarial AI attacks, but intermediate layers where visual and textual information are fused together.

Researchers claim strong black-box transferability

The researchers evaluated the technique against multiple open-source LVLMs, including MiniGPT4, BLIP-2, InstructBLIP, BLIVA, and Qwen2.5-VL, the paper added.

According to the paper, the attack achieved an average success rate of 66.36% across tested models, outperforming prior baseline attacks by roughly 41 percentage points.

The researchers also said the technique demonstrated “strong transferability in black-box settings,” meaning the attacks remained effective even without direct access to a target model’s parameters or architecture.

The paper further claimed the perturbations remained visually stealthy while maintaining effectiveness across multiple LVLM architectures.

No effective defense

The researchers evaluated several defense mechanisms designed to neutralize hidden image manipulations, including random resizing, image rotation, JPEG compression, and inference-level safeguards such as SmoothVLM, a specialized defense framework designed to protect Vision-Language Models (VLMs) from patched visual prompt injections, and DPS, which guides models using partial image views.

According to the paper, SmoothVLM proved the most effective, reducing attack success rates to below 5% in several scenarios, while JPEG compression also weakened the attacks by suppressing high-frequency image artifacts.

However, the researchers said none of the tested defenses completely eliminated the attacks, suggesting stronger multimodal AI security protections may still be needed.

Enterprise AI deployments may widen exposure

The research arrives as enterprises rapidly expand deployments of multimodal AI systems capable of processing screenshots, PDFs, dashboards, forms, video streams, and enterprise documents alongside natural language prompt

SentinelOne
Related Product

SentinelOne Control

by SentinelOne$8.00/mo

Prevention and detection of attacks across all major vectors with fully automated, policy-driven response capabilities and real-time forensics.

Learn More
s.

The researchers noted that adversarial examples generated using the technique could potentially “mislead VLM-based web agents” and “disrupt real-world object detectors.”

“Even if textual inputs are sanitized, manipulated images can still subvert the model’s outputs or actions,” Kaushik said.

She said organizations that use multimodal AI for document processing, customer interactions, content moderation, and autonomous systems may face increasing exposure to adversarial image manipulation and prompt injection attacks.

“Security controls designed for unimodal systems are insufficient,” Kaushik said. The researchers acknowledged that the work was conducted in controlled research settings using open-source models and did not describe observed exploitation in real-world enterprise environments.

Gyana Swain is a seasoned technology journalist with over 20 years' experience covering the telecom and IT space. He is a consulting editor with VARINDIA and earlier in his career, he held editorial positions at CyberMedia, PTI, 9dot9 Media, and Dennis Publishing. A published author of two books, he combines industry insight with narrative depth. Outside of work, he’s a keen traveler and cricket enthusiast. He earned a B.S. degree from Utkal University.

Show me more

GitHub admits major source code leak after 3,800 internal repositories breached

SHub Reaper impersonates Apple, Google, and Microsoft in one MacOS attack chain

Microsoft disrupts malware code-signing service used by ransomware gangs

CSO Executive Sessions ASEAN: From Compliance to Cyber Resilience-Securing Patient Trust in Southeast Asia’s Hospitals

How Intelligence and AI Are Changing Cyber Defense | Erin Whitmore, Former CIA

Inside the SMB Threat Landscape: AT&T’s Senthil Ramakrishnan on Why Small Businesses Are Cybercrime’s Favorite Target

The Human Side of Cybersecurity: Stress, Deepfakes & the Hidden Cost of Breaches

CSO Executive Sessions ASEAN: From Compliance to Cyber Resilience-Securing Patient Trust in Southeast Asia’s Hospitals

How Intelligence and AI Are Changing Cyber Defense | Erin Whitmore, Former CIA

Originally published on May 18, 2026 by CSO Online.

Expert Analysis for Orange County Businesses

How this impacts SMBs in Southern California

Tech News Roundup: New image-based prompt injection attack targets multimodal AI models

The recent discovery of the “CrossMPI” technique, which enables attackers to manipulate AI systems through subtle image perturbations, poses significant implications for small and medium businesses (SMBs) in Orange County and Southern California. As companies increasingly integrate multimodal AI into their operations—ranging from customer service chatbots to document processing tools—this vulnerability could lead to serious misinterpretations of user instructions, resulting in operational mistakes or compromised data integrity. Given that 80% of enterprise software is projected to be multimodal by 2030, the urgency for robust security measures is paramount, especially for SMBs that may lack the resources to develop comprehensive defenses.

Addressing these emerging threats is where The Network Company can assist local businesses through tailored cybersecurity solutions and managed IT services. By implementing rigorous monitoring systems, such as our Security Operations Center (SOC) services, we can help detect and respond to potential attacks in real-time. Additionally, our cloud backup and BCDR services ensure that your critical data remains safe and recoverable, even in the event of an attack. To navigate the expanding landscape of AI security risks effectively, partnering with TNC means gaining access to expert insights and proactive measures designed to safeguard your business's digital environment.

Read the full article: Original Source Link

Frequently Asked Questions

What is CrossMPI?

CrossMPI is a technique that uses image perturbations to manipulate how AI models interpret user instructions.

Why is this attack significant?

It poses new security risks as enterprises increasingly adopt multimodal AI systems in their operations.

Content Quality

Last Reviewed:May 20, 2026
Fact-Checked:May 20, 2026
0
Word Count:1,483 words

Article Tags

Artificial Intelligence, Cyberattacks, Generative AI, Security
image-based prompt injection
multimodal AI security
CrossMPI technique
AI model vulnerabilities
generative AI risks
vision-language models
enterprise AI security
cybersecurity threats
M

About Microsoft disrupts malware code-signing service used by ransomware gangs | CSO Online

Content Author

With extensive experience in cybersecurity and AI technologies, the author specializes in analyzing emerging threats to AI systems and providing insights for businesses on safeguarding their multimodal AI applications.

Expertise: Cybersecurity, Artificial Intelligence, Prompt Injection Attacks, Multimodal AI, Risk Management, Generative AI, Threat Analysis

Ready to Transform Your Business?

Discover how our expert IT solutions can help your business thrive.