# multimodal
標記為「multimodal」的 234 篇文章
Advanced Practice Exam
25-question practice exam covering advanced AI red team techniques: multimodal attacks, training pipeline exploitation, agentic system attacks, embedding manipulation, and fine-tuning security.
Practice Exam 2: Advanced AI Security
25-question advanced practice exam covering multimodal attacks, training pipeline security, cloud AI security, forensics, and governance.
Multimodal Security Practice Exam
Practice exam covering image injection, audio attacks, cross-modal transfer, and document parsing exploitation.
Advanced Multimodal Assessment
In-depth assessment of cross-modal attack vectors including image injection, audio manipulation, and steganographic techniques.
Multimodal Attacks Assessment
Assessment on image injection, audio attacks, cross-modal exploitation, and multimodal defense bypass.
Multimodal Defense Assessment
Assessment covering defenses against visual injection, audio attacks, and cross-modal exploitation.
Multimodal Attack Assessment
Test your understanding of attacks against multimodal AI systems, including image-based injection, audio adversarial examples, and cross-modal manipulation with 10 intermediate-level questions.
Advanced Multimodal Attacks Assessment
Advanced assessment covering cross-modal attacks, steganographic injection, and multimodal defense bypass.
Skill Verification: Multimodal Attack Execution
Hands-on verification of image injection, audio manipulation, and cross-modal transfer attacks.
Skill Verification: Multimodal Defense
Practical verification of ability to defend multimodal AI systems against cross-modal attacks.
Skill Verification: Multimodal Attacks
Hands-on verification of multimodal attack capabilities across image, audio, and document modalities.
Multimodal Security Study Guide
Study guide for multimodal attack and defense topics covering image, audio, and document modalities.
Multimodal Security Study Guide (Assessment)
Study guide for multimodal security assessments covering image, audio, document, and cross-modal attacks.
Capstone: Build a Multimodal Attack Testing Suite
Design and implement a comprehensive testing suite for attacking multimodal AI systems across text, image, audio, and document modalities.
Capstone: Multimodal System Assessment
Capstone exercise: red team assessment of a multimodal AI system processing images, documents, and text.
Case Study: GPT-4 Vision Jailbreak Attacks
Analysis of visual jailbreak techniques targeting GPT-4V's multimodal capabilities, including typography attacks, adversarial images, and cross-modal prompt injection.
Case Study: Multimodal Jailbreak Campaigns
Analysis of multimodal jailbreak campaigns targeting GPT-4V and Gemini vision capabilities.
Case Study: Prompt Injection Attacks on Google Bard/Gemini
Analysis of prompt injection vulnerabilities discovered in Google Bard (later Gemini), including indirect injection through Google Workspace integration and the unique attack surface created by multimodal capabilities.
Multimodal Embedding Attacks
Exploiting cross-modal embedding models like CLIP — adversarial image-text alignment manipulation, cross-modal injection, and attacks on multimodal retrieval systems.
Multimodal Embedding Attacks (Embedding Vector Security)
Attacking multimodal embedding spaces like CLIP for cross-modal manipulation.
Model Types and Their Attack Surfaces
How text, vision, multimodal, embedding, and code generation models each present unique vulnerabilities and attack surfaces for red teamers.
Multi-Modal Reasoning Attacks
Attacking reasoning processes that span multiple modalities in vision-language and audio-language models.
Multimodal Reasoning Security
Security challenges specific to models that reason across text, image, audio, and video modalities simultaneously.
Injection Research
Advanced research in prompt injection, jailbreak automation, and multimodal attack vectors, covering cutting-edge techniques that push beyond standard injection approaches.
Adversarial Perturbation Attacks
Gradient-based pixel-level attacks against vision encoders, covering FGSM, PGD, C&W, transferability, physical-world adversarial examples, and perturbation budget constraints.
Audio & Speech Adversarial Attacks
Adversarial attacks against speech-enabled AI systems, covering ultrasonic injection, ASR adversarial noise, hidden voice commands, voice cloning for authentication bypass, and real-time audio manipulation.
Multimodal Attack Vectors
Exploitation of vision-language models, typographic attacks, audio injection, document-based attacks, and cross-modal adversarial techniques.
Cross-Modal Embedding Attacks
Exploitation of shared embedding spaces across modalities: CLIP adversarial images, typographic attacks, VLM injection, and dimensionality reduction attacks.
Lab: Audio Adversarial Examples
Hands-on lab for crafting adversarial audio perturbations that cause speech-to-text models and voice assistants to misinterpret spoken commands, demonstrating attacks on audio AI systems.
Multimodal Attack Chain Lab
Chain attacks across text, image, and structured data modalities to exploit multimodal system vulnerabilities.
Lab: Multimodal Attack Pipeline
Build an automated multimodal attack pipeline that generates adversarial images, combines them with text prompts, and tests against vision-language models (VLMs).
Multimodal Image Injection
Embed adversarial text in images that triggers prompt injection in vision-language models.
Multi-Modal Attack Chain Orchestration
Orchestrate attacks across text, image, and document modalities to bypass per-modality safety filters.
Lab: Multimodal Input Testing Basics
Introduction to testing multimodal LLMs with image and text inputs to understand cross-modal behavior.
CTF: Multimodal Maze
Navigate a multimodal challenge using image, text, and audio injection vectors. Each modality unlocks the next stage of the maze, requiring cross-modal attack chaining.
Multimodal Cipher: Cross-Modal Decryption
Decode a flag split across text, image, and audio inputs processed by a multimodal AI system.
Lab: Multimodal Injection
Hands-on lab for embedding text instructions in images to exploit vision-enabled LLMs. Learn to craft visual prompt injections, test OCR-based attacks, and evaluate multimodal safety boundaries.
Lab: Multimodal Injection (Intermediate Lab)
Embed prompt injection instructions in images for vision-enabled models, testing how visual content can carry adversarial payloads.
Lab: Intermediate Multimodal Security Testing
Test multimodal LLMs with crafted images containing embedded text, adversarial perturbations, and visual injection payloads.
Simulation: Multimodal Application Assessment
Red team simulation targeting an application that processes both images and text, testing visual injection, cross-modal attacks, and multimodal jailbreaks.
Gemini (Google) Overview
Architecture overview of Google's Gemini model family, including natively multimodal design, long context capabilities, Google ecosystem integration, and security-relevant features for red teaming.
Gemini Known Vulnerabilities
Documented Gemini vulnerabilities including image generation bias incidents, system prompt extraction, safety filter inconsistencies, multimodal injection exploits, and grounding abuse.
Multimodal Model Security Comparison
Comparing security properties across multimodal models (GPT-4V, Claude, Gemini) with focus on cross-modal injection and vision-language attacks.
3D Model Adversarial Attacks
Adversarial attacks on AI systems that process 3D models, point clouds, and spatial data.
Adversarial Image Perturbation for VLMs
Generating adversarial perturbations that cause vision-language models to misinterpret or follow injected instructions.
Adversarial Patch Attacks on VLMs
Crafting physical adversarial patches that trigger specific behaviors in vision-language models when captured by cameras.
Adversarial Typography Attacks
Craft adversarial text rendered as images to exploit OCR and vision model text recognition.
Audio Modality Attacks
Comprehensive attack taxonomy for audio-enabled LLMs: adversarial audio generation, voice-based prompt injection, cross-modal split attacks, and ultrasonic perturbations.
Audio Model Attack Surface
Overview of audio model security, including attacks on Whisper, speech-to-text systems, voice assistants, and the audio processing pipeline.
Adversarial Attacks on Audio and Speech Models
Techniques for crafting adversarial audio that exploits speech recognition, voice assistants, and audio-language models including hidden commands and psychoacoustic masking.
Audio Frequency Domain Injection
Hiding adversarial instructions in audio frequency bands that are processed by speech-to-text models but inaudible to humans.
Hidden Audio Commands for Voice AI
Embed hidden commands in audio that are inaudible to humans but recognized by speech processing AI.
Audio-Based Injection Attacks
Attacking speech-to-text and audio-language models through adversarial audio crafting.
Chart and Graph Injection Attacks
Embedding adversarial instructions in charts, graphs, and data visualizations processed by VLMs.
Cross-Modal Attack Strategies
Overview of attack strategies that exploit the boundaries between input modalities in multimodal AI systems, including vision-language, audio-text, and document processing pipelines.
Lab: Multi-Modal Attack Chain
Hands-on lab for building and executing a multi-step attack chain that combines image injection, document exploitation, and text-based techniques against a multimodal AI system.
Multimodal Defense Strategies
Comprehensive defense approaches for multimodal AI systems: cross-modal verification, perceptual hashing, NSFW detection, input sanitization, and defense-in-depth architectures.
Multimodal Jailbreaking Techniques
Combined multi-modal approaches to bypass safety alignment, including image-text combination attacks, typographic jailbreaks, visual chain-of-thought manipulation, and multi-modal crescendo techniques.
Transferring Attacks Across Modalities
Techniques for crafting adversarial inputs that transfer across modalities, using one input channel to attack processing in another, including image-to-text, audio-to-action, and document-to-tool attack chains.
Cross-Modal Transfer Attacks
Attacks that transfer across modalities — using one input modality to attack processing in another.
Depth Map Adversarial Attacks
Adversarial manipulation of depth information in 3D understanding tasks processed by multimodal models.
Attacks on Document Processing AI
Techniques for attacking document understanding systems including OCR pipelines, PDF processors, and document-language models through layout manipulation, hidden text, and metadata injection.
Document Metadata Injection
Inject adversarial content through document metadata fields processed by multimodal AI systems.
Document Parsing Exploitation
Exploiting PDF, DOCX, and other document parsers in multimodal AI systems for injection and data extraction.
Image-Based Prompt Injection Techniques
Techniques for embedding adversarial prompts in images consumed by vision-language models.
Image Metadata Injection Attacks
Exploiting EXIF metadata, IPTC data, and other image metadata fields for prompt injection in VLM pipelines.
Image Steganography for AI Attacks
Using steganographic techniques to embed adversarial payloads in images that evade human inspection and automated detection while influencing AI model behavior.
Image Steganography for LLM Injection
Use image steganography to embed prompt injection payloads invisible to human viewers.
Multimodal Security
Security assessment of multimodal AI systems processing images, audio, video, and cross-modal inputs, covering vision-language models, speech systems, video analysis, and cross-modal attack techniques.
Medical Imaging Adversarial Attacks
Adversarial attacks on medical imaging AI including radiology, pathology, and dermatology classification systems.
Alignment Challenges in Multimodal Models
Analysis of alignment challenges specific to multimodal AI systems, including cross-modal safety gaps, representation conflicts, and the difficulty of extending text-based safety training to visual, audio, and video inputs.
Multimodal Consistency Attacks
Exploit inconsistencies between how different modalities process the same information.
Multimodal Defense Bypass Techniques
Techniques for bypassing safety filters that only analyze individual modalities.
Defending Multimodal AI Systems
Comprehensive defense strategies for multimodal AI systems including input sanitization, cross-modal safety classifiers, instruction hierarchy, and monitoring for adversarial multimodal inputs.
Multimodal Fusion Layer Attacks
Attacking the fusion mechanisms that combine information from multiple modalities in multimodal models.
Model Extraction from Multimodal Systems
Techniques for extracting model capabilities, weights, and architecture details from multimodal AI systems through visual, audio, and cross-modal query strategies.
Image-Based Prompt Injection Attacks
Comprehensive techniques for injecting adversarial prompts through images, covering typographic injection, steganographic embedding, and visual payload delivery against multimodal AI systems.
Multimodal Prompt Injection Survey
Comprehensive survey of prompt injection vectors across all modalities including text, image, audio, video, and code.
Multimodal RAG Poisoning
Poisoning multimodal RAG systems through adversarial documents with embedded visual and textual payloads.
Methodology for Red Teaming Multimodal Systems
Structured methodology for conducting security assessments of multimodal AI systems, covering scoping, attack surface enumeration, test execution, and reporting with MITRE ATLAS mappings.
Benchmarking Multimodal Model Safety
Designing and implementing safety benchmarks for multimodal AI models that process images, audio, and video alongside text, covering cross-modal attack evaluation, consistency testing, and safety score aggregation.
Multimodal Watermark Evasion
Techniques for evading and removing watermarks applied to AI-generated images, audio, and video content.
OCR Adversarial Attacks
Crafting images that cause OCR systems to extract adversarial text for downstream injection.
PDF Document Injection Techniques
Exploiting PDF parsing in multimodal models to inject instructions through hidden text layers and embedded objects.
QR Code and Barcode Injection Attacks
Using QR codes and barcodes as vectors for prompt injection in vision-language model applications.
Satellite Imagery Adversarial Attacks
Adversarial manipulation of satellite imagery analysis AI for geospatial intelligence and earth observation.
Attacks via Screen Capture and Computer-Use AI
Techniques for attacking AI systems that process screen captures, including computer-use agents, screen-reading assistants, and automated UI testing systems.
Screenshot and UI Injection Attacks
Injecting prompts through screenshots and UI elements processed by computer-use AI agents.
Sign Language and Gesture Injection
Exploiting sign language and gesture recognition models through adversarial physical gestures and modified inputs.
Steganographic Prompt Injection
Hiding prompt injection payloads using steganographic techniques in images and audio.
Adversarial Attacks on Text-to-Image Models
Understanding and evaluating adversarial attacks on text-to-image generation models including prompt manipulation for safety bypass, concept erasure attacks, adversarial perturbation of guidance, and membership inference on training data.
Typography-Based Prompt Injection
Exploiting text rendering in images to deliver prompt injection payloads through typography recognition in VLMs.
Video Model Attacks
Video understanding model security, frame-level vs temporal attacks, how video models process sequences, and the complete attack surface overview.
Video Understanding Model Exploitation
Attacking video captioning, video Q&A, and action recognition models with adversarial videos that cause misclassification or instruction injection.
Video Frame Injection
Injecting adversarial content into video frames processed by video-understanding AI models.
Video Temporal Frame Injection
Injecting adversarial frames at specific temporal positions in video streams processed by video understanding models.
Attacks on Video Understanding Models
Techniques for attacking AI video understanding systems through frame injection, temporal manipulation, and adversarial video generation targeting models like Gemini 2.5 Pro.
Attacks on Vision-Language Models
Comprehensive techniques for attacking vision-language models including GPT-4V, Claude vision, and Gemini, covering adversarial images, typographic exploits, and multimodal jailbreaks.
VLM Architecture & Vision-Language Alignment
Deep dive into VLM architectures including CLIP, SigLIP, and vision transformers. How image patches become tokens, alignment training, and where misalignment creates exploitable gaps.
Image-Based Prompt Injection
Techniques for embedding text instructions in images to manipulate VLMs, including steganographic injection, visible text attacks, and QR code exploitation.
Vision-Language Model Attacks
Comprehensive overview of the VLM attack surface, how vision encoders connect to language models, and why multimodal systems create new injection vectors.
VLM-Specific Jailbreaking
Jailbreaking techniques that exploit the vision modality, including image-text inconsistency attacks, visual safety bypass, and cross-modal jailbreaking strategies.
Multimodal Text Injection Vectors
Injecting adversarial text instructions through non-text modalities including images with embedded text, audio transcription, and document parsing.
Audio Prompt Injection
Injecting adversarial instructions through audio inputs to speech-to-text and multimodal models, exploiting the audio channel as an alternative injection vector.
Cross-Modal Confusion
Confusing multimodal AI models by sending conflicting or complementary signals across different input modalities to bypass safety mechanisms and exploit fusion weaknesses.
Image-Based Prompt Injection (Attack Walkthrough)
Embedding text instructions in images that vision models read, enabling prompt injection through the visual modality to bypass text-only input filters and safety mechanisms.
Multi-Image Chaining
Chaining prompt injection payloads across multiple images in a conversation to deliver complex attacks that evade per-image content filters and build injection context progressively.
Multi-Modal Document Attack Walkthrough
Combine visual and textual injection in documents processed by multimodal LLM applications.
Multimodal Image Injection Walkthrough
Step-by-step walkthrough of embedding adversarial prompts in images for vision model exploitation.
OCR-Based Attacks
Exploiting Optical Character Recognition processing pipelines to inject adversarial text into AI systems, targeting the gap between what OCR extracts and what humans see.
PDF Document Injection
Injecting adversarial prompts through PDF documents processed by AI systems, exploiting document parsing pipelines to deliver payloads through text layers, metadata, and embedded objects.
QR Code Injection
Using QR codes as prompt injection vectors against vision-language models, encoding adversarial instructions in machine-readable formats that models decode and follow.
Steganographic Payload Delivery
Hiding prompt injection payloads in images using steganographic techniques, delivering adversarial instructions through pixel-level modifications invisible to human observers.
Typography Injection in Images
Using rendered text with specific fonts, styles, and typographic techniques in images to inject prompts into vision-language models while evading detection.
Video Frame Injection (Attack Walkthrough)
Embedding prompt injection payloads in specific video frames to attack multimodal models that process video content, exploiting temporal and visual channels simultaneously.
Vision Model Attack Walkthrough
Attack vision-language models through adversarial images with embedded text, perturbations, and visual trojans.
Image Steganography Injection Walkthrough
Walkthrough of hiding prompt injection payloads in images using steganographic techniques for vision model attacks.
進階 Practice Exam
25-question practice exam covering advanced AI red team techniques: multimodal attacks, training pipeline exploitation, agentic system attacks, embedding manipulation, and fine-tuning security.
Practice Exam 2: 進階 AI 安全
25-question advanced practice exam covering multimodal attacks, training pipeline security, cloud AI security, forensics, and governance.
Multimodal 安全 Practice Exam
Practice exam covering image injection, audio attacks, cross-modal transfer, and document parsing exploitation.
進階 Multimodal 評量
In-depth assessment of cross-modal attack vectors including image injection, audio manipulation, and steganographic techniques.
Multimodal 攻擊s 評量
評量 on image injection, audio attacks, cross-modal exploitation, and multimodal defense bypass.
Multimodal 防禦 評量
評量 covering defenses against visual injection, audio attacks, and cross-modal exploitation.
章節評量：多模態安全
15 題校準評量，測試你對多模態 AI 安全的理解——視覺注入、音訊攻擊、跨模態利用。
進階 Multimodal 攻擊s 評量
進階 assessment covering cross-modal attacks, steganographic injection, and multimodal defense bypass.
Skill Verification: Multimodal 攻擊 Execution
Hands-on verification of image injection, audio manipulation, and cross-modal transfer attacks.
Skill Verification: Multimodal 防禦
Practical verification of ability to defend multimodal AI systems against cross-modal attacks.
Skill Verification: Multimodal 攻擊s
Hands-on verification of multimodal attack capabilities across image, audio, and document modalities.
Multimodal 安全 Study 指南
Study guide for multimodal attack and defense topics covering image, audio, and document modalities.
Multimodal 安全 Study 指南 (評量)
Study guide for multimodal security assessments covering image, audio, document, and cross-modal attacks.
多模態攻擊版圖
隨著 AI 系統處理圖片、音訊與影片以及文字，攻擊面已大幅擴展。紅隊員需要知道的事。
Capstone: Build a Multimodal 攻擊 Testing Suite
Design and implement a comprehensive testing suite for attacking multimodal AI systems across text, image, audio, and document modalities.
Capstone: Multimodal System 評量
Capstone exercise: red team assessment of a multimodal AI system processing images, documents, and text.
Case Study: GPT-4 Vision 越獄 攻擊s
Analysis of visual jailbreak techniques targeting GPT-4V's multimodal capabilities, including typography attacks, adversarial images, and cross-modal prompt injection.
Case Study: Multimodal 越獄 Campaigns
Analysis of multimodal jailbreak campaigns targeting GPT-4V and Gemini vision capabilities.
Case Study: 提示詞注入 攻擊s on Google Bard/Gemini
Analysis of prompt injection vulnerabilities discovered in Google Bard (later Gemini), including indirect injection through Google Workspace integration and the unique attack surface created by multimodal capabilities.
多模態嵌入向量攻擊
利用 CLIP 等跨模態嵌入模型——對抗性圖文對齊操控、跨模態注入與對多模態檢索系統的攻擊。
Multimodal Embedding 攻擊s (Embedding Vector 安全)
攻擊ing multimodal embedding spaces like CLIP for cross-modal manipulation.
模型類型與其攻擊面
文字、視覺、多模態、embedding 與程式碼生成模型如何各呈現紅隊員獨特之漏洞與攻擊面。
Multi-Modal Reasoning 攻擊s
攻擊ing reasoning processes that span multiple modalities in vision-language and audio-language models.
Multimodal Reasoning 安全
安全 challenges specific to models that reason across text, image, audio, and video modalities simultaneously.
注入研究
提示詞注入、越獄自動化與多模態攻擊向量的進階研究，涵蓋超越標準注入方法的尖端技術。
Adversarial Perturbation 攻擊s
Gradient-based pixel-level attacks against vision encoders, covering FGSM, PGD, C&W, transferability, physical-world adversarial examples, and perturbation budget constraints.
Audio & Speech Adversarial 攻擊s
Adversarial attacks against speech-enabled AI systems, covering ultrasonic injection, ASR adversarial noise, hidden voice commands, voice cloning for authentication bypass, and real-time audio manipulation.
Multimodal 攻擊 Vectors
利用ation of vision-language models, typographic attacks, audio injection, document-based attacks, and cross-modal adversarial techniques.
跨模態 Embedding 攻擊
跨模態之共享 embedding 空間利用：CLIP 對抗影像、印刷式攻擊、VLM 注入與維度縮減攻擊。
實驗室: Audio Adversarial Examples
Hands-on lab for crafting adversarial audio perturbations that cause speech-to-text models and voice assistants to misinterpret spoken commands, demonstrating attacks on audio AI systems.
Multimodal 攻擊 Chain 實驗室
Chain attacks across text, image, and structured data modalities to exploit multimodal system vulnerabilities.
實驗室: Multimodal 攻擊 Pipeline
Build an automated multimodal attack pipeline that generates adversarial images, combines them with text prompts, and tests against vision-language models (VLMs).
Multimodal Image Injection
Embed adversarial text in images that triggers prompt injection in vision-language models.
Multi-Modal 攻擊 Chain Orchestration
Orchestrate attacks across text, image, and document modalities to bypass per-modality safety filters.
實驗室: Multimodal Input Testing Basics
介紹 to testing multimodal LLMs with image and text inputs to understand cross-modal behavior.
CTF: Multimodal Maze
Navigate a multimodal challenge using image, text, and audio injection vectors. Each modality unlocks the next stage of the maze, requiring cross-modal attack chaining.
Multimodal Cipher: Cross-Modal Decryption
Decode a flag split across text, image, and audio inputs processed by a multimodal AI system.
實驗室: Multimodal Injection
Hands-on lab for embedding text instructions in images to exploit vision-enabled LLMs. Learn to craft visual prompt injections, test OCR-based attacks, and evaluate multimodal safety boundaries.
實驗室: Multimodal Injection (中階 實驗室)
Embed prompt injection instructions in images for vision-enabled models, testing how visual content can carry adversarial payloads.
實驗室: 中階 Multimodal 安全 Testing
Test multimodal LLMs with crafted images containing embedded text, adversarial perturbations, and visual injection payloads.
模擬：多模態應用程式評估
針對同時處理圖片與文字之應用程式的紅隊模擬，測試視覺注入、跨模態攻擊與多模態越獄。
Gemini（Google）概觀
Google Gemini 模型家族的架構概觀，包括原生多模態設計、長上下文能力、Google 生態整合，以及對紅隊具意義的安全相關特性。
Gemini 已知漏洞
已記錄的 Gemini 漏洞，包括影像生成偏見事件、系統提示擷取、安全過濾不一致、多模態注入 exploit，以及 grounding 濫用。
Multimodal 模型 安全 Comparison
Comparing security properties across multimodal models (GPT-4V, Claude, Gemini) with focus on cross-modal injection and vision-language attacks.
3D 模型 Adversarial 攻擊s
Adversarial attacks on AI systems that process 3D models, point clouds, and spatial data.
Adversarial Image Perturbation for VLMs
Generating adversarial perturbations that cause vision-language models to misinterpret or follow injected instructions.
Adversarial Patch 攻擊s on VLMs
Crafting physical adversarial patches that trigger specific behaviors in vision-language models when captured by cameras.
Adversarial Typography 攻擊s
Craft adversarial text rendered as images to exploit OCR and vision model text recognition.
Audio Modality 攻擊s
Comprehensive attack taxonomy for audio-enabled LLMs: adversarial audio generation, voice-based prompt injection, cross-modal split attacks, and ultrasonic perturbations.
Adversarial 攻擊s on Audio and Speech 模型s
Techniques for crafting adversarial audio that exploits speech recognition, voice assistants, and audio-language models including hidden commands and psychoacoustic masking.
Audio Frequency Domain Injection
Hiding adversarial instructions in audio frequency bands that are processed by speech-to-text models but inaudible to humans.
Hidden Audio Commands for Voice AI
Embed hidden commands in audio that are inaudible to humans but recognized by speech processing AI.
Audio-Based Injection 攻擊s
攻擊ing speech-to-text and audio-language models through adversarial audio crafting.
Chart and Graph Injection 攻擊s
Embedding adversarial instructions in charts, graphs, and data visualizations processed by VLMs.
跨模態攻擊
跨多個輸入模態串接漏洞的攻擊——涵蓋基於文件的攻擊、多模態越獄、模態橋接與資訊洩漏。
實驗室: Multi-Modal 攻擊 Chain
Hands-on lab for building and executing a multi-step attack chain that combines image injection, document exploitation, and text-based techniques against a multimodal AI system.
Multimodal 防禦 Strategies
Comprehensive defense approaches for multimodal AI systems: cross-modal verification, perceptual hashing, NSFW detection, input sanitization, and defense-in-depth architectures.
多模態越獄技術
結合多模態途徑以繞過安全對齊，含圖像－文字組合攻擊、排字越獄、視覺思維鏈操弄，與多模態漸進技術。
Transferring 攻擊s Across Modalities
Techniques for crafting adversarial inputs that transfer across modalities, using one input channel to attack processing in another, including image-to-text, audio-to-action, and document-to-tool attack chains.
Cross-Modal Transfer 攻擊s
攻擊s that transfer across modalities — using one input modality to attack processing in another.
Depth Map Adversarial 攻擊s
Adversarial manipulation of depth information in 3D understanding tasks processed by multimodal models.
攻擊s on Document Processing AI
Techniques for attacking document understanding systems including OCR pipelines, PDF processors, and document-language models through layout manipulation, hidden text, and metadata injection.
Document Metadata Injection
Inject adversarial content through document metadata fields processed by multimodal AI systems.
Document Parsing 利用ation
利用ing PDF, DOCX, and other document parsers in multimodal AI systems for injection and data extraction.
Image-Based 提示詞注入 Techniques
Techniques for embedding adversarial prompts in images consumed by vision-language models.
Image Metadata Injection 攻擊s
利用ing EXIF metadata, IPTC data, and other image metadata fields for prompt injection in VLM pipelines.
Image Steganography for AI 攻擊s
Using steganographic techniques to embed adversarial payloads in images that evade human inspection and automated detection while influencing AI model behavior.
Image Steganography for LLM Injection
Use image steganography to embed prompt injection payloads invisible to human viewers.
多模態安全
處理影像、音訊、影片與跨模態輸入之多模態 AI 系統的安全評估，涵蓋視覺-語言模型、語音系統、影片分析與跨模態攻擊技術。
Medical Imaging Adversarial 攻擊s
Adversarial attacks on medical imaging AI including radiology, pathology, and dermatology classification systems.
Alignment Challenges in Multimodal 模型s
Analysis of alignment challenges specific to multimodal AI systems, including cross-modal safety gaps, representation conflicts, and the difficulty of extending text-based safety training to visual, audio, and video inputs.
Multimodal Consistency 攻擊s
利用 inconsistencies between how different modalities process the same information.
Multimodal 防禦 Bypass Techniques
Techniques for bypassing safety filters that only analyze individual modalities.
Defending Multimodal AI Systems
Comprehensive defense strategies for multimodal AI systems including input sanitization, cross-modal safety classifiers, instruction hierarchy, and monitoring for adversarial multimodal inputs.
Multimodal Fusion Layer 攻擊s
攻擊ing the fusion mechanisms that combine information from multiple modalities in multimodal models.
模型 Extraction from Multimodal Systems
Techniques for extracting model capabilities, weights, and architecture details from multimodal AI systems through visual, audio, and cross-modal query strategies.
Image-Based 提示詞注入 攻擊s
Comprehensive techniques for injecting adversarial prompts through images, covering typographic injection, steganographic embedding, and visual payload delivery against multimodal AI systems.
Multimodal 提示詞注入 Survey
Comprehensive survey of prompt injection vectors across all modalities including text, image, audio, video, and code.
Multimodal RAG 投毒
投毒 multimodal RAG systems through adversarial documents with embedded visual and textual payloads.
Methodology for 紅隊演練 Multimodal Systems
Structured methodology for conducting security assessments of multimodal AI systems, covering scoping, attack surface enumeration, test execution, and reporting with MITRE ATLAS mappings.
Benchmarking Multimodal 模型 Safety
Designing and implementing safety benchmarks for multimodal AI models that process images, audio, and video alongside text, covering cross-modal attack evaluation, consistency testing, and safety score aggregation.
Multimodal Watermark Evasion
Techniques for evading and removing watermarks applied to AI-generated images, audio, and video content.
OCR Adversarial 攻擊s
Crafting images that cause OCR systems to extract adversarial text for downstream injection.
PDF Document Injection Techniques
利用ing PDF parsing in multimodal models to inject instructions through hidden text layers and embedded objects.
QR Code and Barcode Injection 攻擊s
Using QR codes and barcodes as vectors for prompt injection in vision-language model applications.
Satellite Imagery Adversarial 攻擊s
Adversarial manipulation of satellite imagery analysis AI for geospatial intelligence and earth observation.
攻擊s via Screen Capture and Computer-Use AI
Techniques for attacking AI systems that process screen captures, including computer-use agents, screen-reading assistants, and automated UI testing systems.
Screenshot and UI Injection 攻擊s
Injecting prompts through screenshots and UI elements processed by computer-use AI agents.
Sign Language and Gesture Injection
利用ing sign language and gesture recognition models through adversarial physical gestures and modified inputs.
Steganographic 提示詞注入
Hiding prompt injection payloads using steganographic techniques in images and audio.
Adversarial 攻擊s on Text-to-Image 模型s
Understanding and evaluating adversarial attacks on text-to-image generation models including prompt manipulation for safety bypass, concept erasure attacks, adversarial perturbation of guidance, and membership inference on training data.
Typography-Based 提示詞注入
利用ing text rendering in images to deliver prompt injection payloads through typography recognition in VLMs.
影片模型攻擊
影片理解模型的安全性、影格層級攻擊與時序攻擊的差異、影片模型如何處理序列資訊，以及完整的攻擊面概觀。
視訊理解模型利用
以造成誤分類或指令注入之對抗視訊攻擊視訊字幕、視訊 Q&A 與動作辨識模型。
Video Frame Injection
Injecting adversarial content into video frames processed by video-understanding AI models.
Video Temporal Frame Injection
Injecting adversarial frames at specific temporal positions in video streams processed by video understanding models.
攻擊s on Video Understanding 模型s
Techniques for attacking AI video understanding systems through frame injection, temporal manipulation, and adversarial video generation targeting models like Gemini 2.5 Pro.
攻擊s on Vision-Language 模型s
Comprehensive techniques for attacking vision-language models including GPT-4V, Claude vision, and Gemini, covering adversarial images, typographic exploits, and multimodal jailbreaks.
VLM 架構與視覺—語言對齊
深入探討 VLM 架構，包括 CLIP、SigLIP 與 vision transformers。圖像 patch 如何變成 token、對齊訓練，以及錯位（misalignment）如何製造可利用之缺口。
以圖像為本之提示注入
將文字指令嵌入圖像以操弄 VLM 之技術，含隱寫注入、可見文字攻擊與 QR 碼利用。
視覺-語言模型
視覺-語言模型（VLM）的安全評估——涵蓋 VLM 架構、圖片注入技術、OCR 與字型攻擊、對抗性圖片生成與 VLM 特定越獄。
VLM 特有的越獄手法
利用視覺模態的越獄技術，包括影像─文字不一致攻擊、視覺安全繞過，以及跨模態越獄策略。
Multimodal Text Injection Vectors
Injecting adversarial text instructions through non-text modalities including images with embedded text, audio transcription, and document parsing.
Audio 提示詞注入
Injecting adversarial instructions through audio inputs to speech-to-text and multimodal models, exploiting the audio channel as an alternative injection vector.
Cross-Modal Confusion
Confusing multimodal AI models by sending conflicting or complementary signals across different input modalities to bypass safety mechanisms and exploit fusion weaknesses.
Image-Based 提示詞注入 (攻擊 導覽)
Embedding text instructions in images that vision models read, enabling prompt injection through the visual modality to bypass text-only input filters and safety mechanisms.
Multi-Image Chaining
Chaining prompt injection payloads across multiple images in a conversation to deliver complex attacks that evade per-image content filters and build injection context progressively.
Multi-Modal Document 攻擊 導覽
Combine visual and textual injection in documents processed by multimodal LLM applications.
Multimodal Image Injection 導覽
Step-by-step walkthrough of embedding adversarial prompts in images for vision model exploitation.
OCR-Based 攻擊s
利用ing Optical Character Recognition processing pipelines to inject adversarial text into AI systems, targeting the gap between what OCR extracts and what humans see.
PDF Document Injection
Injecting adversarial prompts through PDF documents processed by AI systems, exploiting document parsing pipelines to deliver payloads through text layers, metadata, and embedded objects.
QR Code Injection
Using QR codes as prompt injection vectors against vision-language models, encoding adversarial instructions in machine-readable formats that models decode and follow.
Steganographic Payload Delivery
Hiding prompt injection payloads in images using steganographic techniques, delivering adversarial instructions through pixel-level modifications invisible to human observers.
Typography Injection in Images
Using rendered text with specific fonts, styles, and typographic techniques in images to inject prompts into vision-language models while evading detection.
Video Frame Injection (攻擊 導覽)
Embedding prompt injection payloads in specific video frames to attack multimodal models that process video content, exploiting temporal and visual channels simultaneously.
Vision 模型 攻擊 導覽
攻擊 vision-language models through adversarial images with embedded text, perturbations, and visual trojans.
Image Steganography Injection 導覽
導覽 of hiding prompt injection payloads in images using steganographic techniques for vision model attacks.