Adversarial Robustness and Attack Vectors
Research on improving LLM resilience against various attack vectors and understanding vulnerabilities
Adversarial Robustness and Attack Vectors
Research on Large Language Models in Adversarial Robustness and Attack Vectors
Bypassing AI Defenses with No Prior Knowledge
Using CLIP as a surrogate model for no-box adversarial attacks
GazeCLIP: The Future of Gaze Tracking
Enhancing accuracy through text-guided multimodal learning
LLM-Powered Phishing: A New Threat Landscape
Comparing AI-generated vs. human-crafted lateral phishing attacks
Secure Enterprise LLM Platform
Making customized language models accessible while maintaining security
Safer Robot Decision-Making
Using LLM Uncertainty to Enhance Robot Safety and Reliability
WildfireGPT: Intelligent Multi-Agent System for Natural Hazards
Enhancing disaster response with specialized RAG-based LLM systems
Enhanced Security Through Smarter Models
Leveraging Finetuned LLMs as Powerful OOD Detectors
Defending LLMs Against Feedback Manipulation
Robust algorithms for protecting AI systems from adversarial feedback
Cross-Lingual Backdoor Attacks in LLMs
Revealing Critical Security Vulnerabilities Across Languages
Flatter Models, Stronger Defense
Linking Loss Surface Geometry to Adversarial Robustness
Securing LLMs Against Adversarial Attacks
Novel defense strategy using residual stream activation analysis
Discovering Hidden LLM Vulnerabilities
A new approach to identifying realistic toxic prompts that bypass AI safety systems
Fingerprinting LLMs: A New Security Challenge
Identifying specific LLMs with just 8 carefully crafted queries
Defeating Adversarial Phishing Attacks
Evaluating and improving ML-based detection systems against sophisticated threats
Exposing LLM Vulnerabilities
A Novel Approach to Red-teaming for Toxic Content Generation
Defending AI Against Harmful Fine-tuning
Introducing Booster: A Novel Defense for LLM Safety
Advancing Model Extraction Attacks on LLMs
Locality Reinforced Distillation improves attack effectiveness by 11-25%
Securing DNA Language Models Against Attacks
First Comprehensive Assessment of Adversarial Robustness in DNA Classification
Backdoor Threats to Vision-Language Models
Identifying security risks with out-of-distribution data
Bypassing AI Defenses: Smarter Adversarial Attacks
New semantically-consistent approach achieves 96.5% attack success rate
Securing LLM-based Agents
A new benchmark for agent security vulnerabilities and defenses
Exposing VLM Vulnerabilities
Self-supervised adversarial attacks on vision-language models
Security Vulnerabilities in SSM Models
Clean-Label Poisoning Can Undermine Generalization
Hidden Dangers in LLM Alignment
Advanced Backdoor Attacks That Evade Detection
The Dark Side of Web-Connected AI
Emerging security threats from LLMs with internet access
Enhancing YOLO with Contextual Intelligence
How Retriever-Dictionary modules expand object detection beyond single images
The Multilingual Vulnerability Gap
How fine-tuning attacks exploit language diversity in LLMs
Hidden Costs of Faster AI
How acceleration techniques affect bias in LLMs
Unmasking Backdoor Attacks in LLMs
Using AI-generated explanations to detect and understand security vulnerabilities
Vulnerabilities in AI-Powered Robots
Critical security risks in Vision-Language-Action robotics systems
Targeted Bit-Flip Attacks on LLMs
How evolutionary optimization can compromise model security with minimal effort
Fortifying Visual AI Against Attacks
Novel Adversarial Prompt Distillation for Stronger Vision-Language Models
Secure AI Collaboration at the Edge
Building Resilient Multi-Task Language Models Against Adversarial Threats
Hidden Threats in Code Comprehension
How imperceptible code manipulations can deceive AI while fooling humans
Exposing Weaknesses in Time Series LLMs
Uncovering critical security vulnerabilities in forecasting models
Exploiting the Reasoning Vulnerability of LLMs
How the SEED attack compromises LLM safety through subtle error injection
Exposing LLM Vulnerabilities: The AutoDoS Attack
A new black-box approach to force resource exhaustion in language models
Defending LLMs Against Input Attacks
Making Prompt Engineering Robust to Real-World Text Imperfections
Hidden Threats in Language Models
Cross-lingual backdoor attacks that evade detection
The Engorgio Attack: A New LLM Security Threat
How malicious prompts can overwhelm language models
Adaptive Security for LLMs
A New Framework That Balances Security and Usability
Fortifying Vision-Language Models Against Attacks
A Two-Stage Defense Strategy for Visual AI Security
Defending Against LLM Jailbreaking
A Novel Defense Mechanism for Safer AI Systems
The Deception Risk in AI Search Systems
How content injection attacks manipulate search results and AI judges
Boosting LLM Defense Without Retraining
How more compute time creates stronger shields against adversarial attacks
Exposing LLM Vulnerabilities
Why current defenses fail under worst-case attacks
Strategic Information Handling in LLMs
How LLMs reveal, conceal and infer information in competitive scenarios
Rethinking LLM Security Evaluations
Current assessments fail to capture real-world cybersecurity risks
Securing Federated Learning Against Attacks
A Communication-Efficient Approach for Byzantine-Resilient Optimization
Measuring Safety Depth in LLMs
A mathematical framework for robust AI safety guardrails
Exploiting Safety Vulnerabilities in DeepSeek LLM
How fine-tuning attacks can bypass safety mechanisms in Chain-of-Thought models
Exploiting Human Biases in AI Recommendations
How cognitive biases create security vulnerabilities in LLM recommenders
Backdoor Vulnerabilities in AI Vision Systems
Detecting poisoned samples in CLIP models with 98% accuracy
The Distraction Problem in AI
How irrelevant context compromises LLM security
Protecting Medical AI from Theft
Novel attacks expose vulnerabilities in medical imaging models
Navigating the LLM Security Battlefield
Comprehensive Analysis of Adversarial Attacks on Large Language Models
Synthetic Data for Better AI Security
Using LLMs to Generate OOD Data for Robust Classification
Confidence Elicitation: A New LLM Vulnerability
How attackers can extract sensitive information without model access
Hidden Dangers in LLMs
Mapping the Growing Backdoor Threat Landscape
Visual Illusion: The New Frontier in CAPTCHA Security
Combating LLM-powered attacks with human visual perception advantages
Bypassing AI Safety Guardrails
How simple activation shifting can compromise LLM alignment
Rethinking Adversarial Alignment for LLMs
Why current approaches to LLM security fall short
Breaking the Fortress of Language Models
A novel backdoor attack targeting o1-like LLMs' reasoning capabilities
UniGuardian: Unified Defense for LLM Security
A novel approach to detect and prevent multiple types of prompt-based attacks
Defending Against LLM Permutation Attacks
How reordering demonstrations can compromise model security
EigenShield: Fortifying Vision-Language Models
A Novel Defense Against Adversarial Attacks Using Random Matrix Theory
Defending AI Models from Poisoned Training Data
A novel adversarial training approach to counter label poisoning attacks
Breaking LLM Guardrails: Advanced Adversarial Attacks
New semantic objective approach improves jailbreak success by 16%
Emoji-Based Attacks on Language Models
Invisible Vulnerabilities in Modern NLP Systems
Defending AI Against Adversarial Attacks
A robust zero-shot classification approach using CLIP purification
Combating DoS Attacks in LLMs
Detecting and preventing harmful recursive loops in language models
Strengthening LLM Security Through Robustness Testing
New framework detects vulnerabilities in LLM-based NLP applications
Strengthening LLM Robustness Against Prompt Variations
A latent adversarial framework that improves resilience to paraphrased prompts
Testing LLMs Against Adversarial Defenses
Evaluating AI's ability to autonomously exploit security measures
Hijacking LLM Agent Reasoning
A Novel Framework for Comprehensive Security Testing of AI Agents
LLM Safety and Output Length
How longer responses affect model security under adversarial attacks
Security Vulnerabilities in RLHF Platforms
How adversaries can misalign language models through manipulation of reinforcement learning systems
Securing the Gatekeepers: LLM Router Vulnerabilities
First comprehensive security analysis of LLM routing systems across their entire lifecycle
The Repeated Token Vulnerability in LLMs
Understanding and resolving a critical security flaw in language models
Breaking Black-Box AI Models
A simple attack approach achieving over 90% success rate against GPT-4.5/4o/o1
Defeating Face-Morphing Attacks with AI
Zero-Shot Detection Using Multi-Modal LLMs and Vision Models
Boosting Vision-Language Model Security
Evolution-based Adversarial Prompts for Robust AI Systems
Defending LLMs Against Manipulative Attacks
A Temporal Context Awareness Framework for Multi-Turn Security
Autonomous Defense: Next-Gen LLM Security Testing
AI that continuously evolves to find LLM vulnerabilities
The Achilles' Heel of AI Reasoning
How Manipulated Endings Can Override Correct Reasoning in LLMs
Teleporting Security Across Language Models
Zero-shot mitigation of Trojans in LLMs without model-specific alignment data
The Hidden Fragility of LLMs
Understanding and mitigating performance collapse during deployment
Guiding AI Reasoning Through Intervention
A novel approach for controlling LLM behavior during the reasoning process
Defending AI Systems Against Adversarial Attacks
A Universal Detection Framework Using Pre-trained Encoders
Advanced Multi-Turn Red Teaming for LLM Security
Emulating sophisticated adversarial attacks through dual-level learning
Defending Recommender Systems from Attacks
A robust retrieval-augmented framework to combat LLM vulnerabilities
Uncovering LLM Vulnerabilities
New methods to identify and address stability issues in language models
Advancing Face Anti-Spoofing Security
Novel Content-Aware Composite Prompt Engineering for Cross-Domain Protection
VRAG: Smart Defense Against Visual Attacks
Training-Free Detection of Visual Adversarial Patches
Hidden Threats in LLM Recommendation Systems
How adversaries can manipulate rankings while evading detection
Fortifying AI Reward Systems Against Attacks
Adversarial training for more robust AI alignment
Defending Against LLM-Powered Attacks on Rumor Detection
A novel approach to secure social media analysis from AI-generated manipulation
Securing LLMs Against Hidden Threats
Using Influence Functions to Detect Poisoned Fine-tuning Data
LLM Vulnerabilities in Spam Detection
Security weaknesses in AI-powered spam filters
Evaluating LLM-Powered Security Attacks
A critical assessment of benchmarking practices in offensive security
Backdoor Vulnerabilities in LLM Recommendations
Exposing & defending against security threats in LLM-powered recommendation systems
