Jailbreaking Attacks and Defense Mechanisms
Research exploring vulnerabilities in LLMs through jailbreaking attacks and developing effective defense strategies
Jailbreaking Attacks and Defense Mechanisms
Research on Large Language Models in Jailbreaking Attacks and Defense Mechanisms
Certified Defense Against LLM Attacks
First framework providing guaranteed safety against adversarial prompts
Defending Against Backdoor Attacks in Black-Box LLMs
Using Defensive Demonstrations at Test-Time
JailGuard: Defending Against Prompt-Based Attacks
A Universal Framework to Detect Jailbreaking and Hijacking Attempts
Cracking the Safeguards: Understanding LLM Security Vulnerabilities
A representation engineering approach to jailbreaking attacks
Evaluating Jailbreak Attacks on LLMs
A new framework to assess attack effectiveness rather than just model robustness
Hidden Attacks: The New Threat to AI Safety
Exploiting embedding spaces to bypass safety measures in open-source LLMs
Efficient Attacks on LLM Defenses
Making adversarial attacks 1000x more computationally efficient
The Security Paradox in Advanced LLMs
How stronger reasoning abilities create new vulnerabilities
Securing LLMs Against Jailbreak Attacks
A Novel Defense Strategy Without Fine-Tuning
ImgTrojan: The Visual Backdoor Threat
How a single poisoned image can bypass VLM safety barriers
LLM Judge Systems Under Attack
How JudgeDeceiver Successfully Manipulates AI Evaluation Systems
The Crescendo Attack
A New Multi-Turn Strategy for Bypassing LLM Safety Guardrails
The Myth of Trigger Transferability
Challenging assumptions about adversarial attacks across language models
Enhancing AI Safety: The MoTE Framework
Combining reasoning chains with expert mixtures for better LLM alignment
Accelerating Jailbreak Attacks
How Momentum Optimization Makes LLM Security Vulnerabilities More Exploitable
Breaking Through LLM Defenses
New Optimization Method Reveals Security Vulnerabilities in AI Systems
Supercharging LLM Security Testing
A novel approach for discovering diverse attack vectors
SelfDefend: A Practical Shield for LLMs
Empowering LLMs to protect themselves against diverse jailbreak attacks
Structural Vulnerabilities in LLMs
How uncommon text structures enable powerful jailbreak attacks
Combating Jailbreak Attacks on LLMs
A standardized toolkit for evaluating harmful content generation
Rethinking LLM Jailbreaks
Distinguishing True Security Breaches from Hallucinations
Jailbreaking LLMs: Exploiting Alignment Vulnerabilities
A novel attack method that bypasses security measures in language models
Breaking the Logic Chains in LLMs
A Mathematical Framework for Understanding Rule Subversion
Smart Query Refinement for Safer LLMs
Using Reinforcement Learning to Improve Prompts and Prevent Jailbreaks
Breaking Through LLM Guardrails with SeqAR
Uncovering security vulnerabilities in large language models through sequential character prompting
Exposing Biases in LLMs Through Adversarial Testing
How jailbreak prompts reveal hidden biases in seemingly safe models
Surgical Precision for Safer LLMs
Enhancing AI safety through targeted parameter editing
The Analytical Jailbreak Threat
How LLMs' Reasoning Capabilities Create Security Vulnerabilities
Fortifying LLMs Against Tampering
Developing tamper-resistant safeguards for open-weight language models
Breaking Through LLM Defenses
A new framework for systematically testing AI safety filters
Hidden Threats: The 'Carrier Article' Attack
How sophisticated jailbreak attacks can bypass LLM safety guardrails
Testing the Guardians: LLM Security Coverage
New metrics for detecting jailbreak vulnerabilities in LLMs
Unlocking LLM Security Architecture
Identifying critical 'safety layers' that protect aligned language models
Defending LLMs Against Unsafe Feedback
Securing RLHF systems from harmful manipulation
Breaking the Guardians: LLM Jailbreak Attacks
A new efficient method to test LLM security defenses
Defending LLMs Against Adversarial Attacks
Refusal Feature Adversarial Training (ReFAT) for Enhanced Safety
Fortifying LLMs Against Jailbreaking Attacks
A data curation approach to enhance model security during customization
Jailbreak Antidote: Smart Defense for LLMs
Balancing safety and utility with minimal performance impact
The Hidden Danger in LLM Safety Mechanisms
How attackers can weaponize false positives in AI safeguards
Breaking the Guards: Advancing Jailbreak Attacks on LLMs
Novel optimization method achieves 20-30% improvement in bypassing safety measures
Securing LLM Fine-Tuning
A Novel Approach to Mitigate Security Risks in Instruction-Tuned Models
Securing LLMs at the Root Level
A Novel Decoding-Level Defense Strategy Against Harmful Outputs
T-Vaccine: Safeguarding LLMs Against Harmful Fine-Tuning
A targeted layer-wise defense approach for enhanced safety alignment
Identifying Safety-Critical Neurons in LLMs
Mapping the attention heads responsible for safety guardrails
Smarter LLM Safety Guardrails
Balancing security and utility in domain-specific contexts
Enhancing LLM Security Testing
Self-tuning models for more effective jailbreak detection
Stealth Attacks on AI Guardrails
New jailbreak vulnerability using benign data mirroring
Self-Hacking VLMs: The IDEATOR Approach
Using AI to discover its own security vulnerabilities
Emoji Attack: Undermining AI Safety Guards
How emojis can bypass safety detection systems in Large Language Models
SQL Injection Jailbreak: A New LLM Security Threat
Exploiting structural vulnerabilities in language models
SequentialBreak: Large Language Models Can be Fooled by Embe...
By Bijoy Ahmed Saiem, MD Sadik Hossain Shanto...
Bypassing AI Guardrails with Moralized Deception
Testing how ethical-sounding prompts can trick advanced LLMs into harmful outputs
When Safety Training Falls Short
Testing LLM safety against natural, semantically-related harmful prompts
Guarding LLMs Against Jailbreak Attacks
A proactive defense system that identifies harmful queries before they reach the model
Restoring Safety in Fine-Tuned LLMs
A post-hoc approach to recover safety alignment after fine-tuning
Exploiting Metaphors to Bypass AI Safety
How imaginative language can jailbreak language models
Breaking LLM Safety Guards
A Simple Yet Effective Approach to LLM Jailbreaking
Hidden in Plain Sight: LLM Security Threats
How human-readable adversarial prompts bypass security measures
Defending Against Jailbreak Attacks in LLMs
A novel layer-level approach for enhancing AI safety
Evolving Jailbreak Attacks on LLMs
A more efficient approach through pattern and behavior learning
The Infinite Jailbreak Problem
How advanced LLMs become easier to manipulate through paraphrasing
The Trojan Horse Technique in LLM Security
How harmless story endings can bypass security safeguards
Humor as a Security Shield
Strengthening LLM Defenses Against Injection Attacks
The Scientific Trojan Horse
How scientific language can bypass LLM safety guardrails
Unveiling LLM Safety Mechanisms
Extracting and analyzing safety classifiers to combat jailbreak attacks
Exposing LLM Security Vulnerabilities
A novel approach to understanding and preventing LLM jailbreaking
The Virus Attack: A Critical Security Vulnerability in LLMs
Bypassing Safety Guardrails Through Strategic Fine-tuning
Guarding the Gates: LLM Security Red-Teaming
Detecting and preventing jailbreaking in conversational AI systems
Defending LLMs Against Harmful Fine-tuning
Simple random perturbations outperform complex defenses
Breaking LLM Safeguards with Universal Magic Words
How embedding-based security can be bypassed with simple text patterns
Exploiting LLM Vulnerabilities: The Indiana Jones Method
How inter-model dialogues create nearly perfect jailbreaks
Fortifying AI Defense Systems
Constitutional Classifiers: A New Shield Against Universal Jailbreaks
Proactive Defense Against LLM Jailbreaks
Using Safety Chain-of-Thought (SCoT) to strengthen model security
Blocking LLM Jailbreaks with Smart Defense
Nearly 100% effective method to detect and prevent prompt manipulation attacks
Jailbreaking LLMs at Scale
How Universal Multi-Prompts Improve Attack Efficiency
Exposing the Guardian Shield of AI
A novel technique to detect guardrails in conversational AI systems
Adversarial Reasoning vs. LLM Safeguards
New methodologies to identify and strengthen AI security vulnerabilities
Countering LLM Jailbreak Attacks
A three-pronged defense strategy against many-shot jailbreaking
Rethinking AI Safety with Introspective Reasoning
Moving beyond refusals to safer, more resilient language models
Breaking the Jailbreakers
Enhancing Security Through Attack Transferability Analysis
Defending LLMs Against Jailbreak Attacks
Why shorter adversarial training is surprisingly effective against complex attacks
Everyday Jailbreaks: The Unexpected Security Gap
How simple conversations can bypass LLM safety guardrails
Defending LLMs from Jailbreak Attacks
A novel defense framework based on concept manipulation
When AI Deception Bypasses Safety Guards
How language models can be manipulated to override safety mechanisms
The Jailbreak Paradox
How LLMs Can Become Their Own Security Threats
Exploiting LLM Security Weaknesses
A novel approach to jailbreaking aligned language models
Smarter Jailbreak Attacks on LLMs
Boosting Attack Efficiency with Compliance Refusal Initialization
Reasoning-Enhanced Attacks on LLMs
A novel framework for detecting security vulnerabilities in conversational AI
Hidden Vulnerabilities: The R2J Jailbreak Threat
How rewritten harmful requests evade LLM safety guardrails
Detecting LLM Safety Vulnerabilities
A fine-grained benchmark for multi-turn dialogue safety
Exploiting LLM Vulnerabilities
A new context-coherent jailbreak attack method bypasses safety guardrails
Strengthening LLM Security Against Jailbreaks
A Dynamic Defense Approach with Minimal Performance Impact
Bypassing LLM Safety Guardrails
How structure transformation attacks can compromise even the most secure LLMs
Exploiting Safety Reasoning in LLMs
How chain-of-thought safety mechanisms can be bypassed
Strengthening LLM Defense Against Jailbreaking
Using reasoning abilities to enhance AI safety
Defending Against LLM Jailbreaks
ShieldLearner: A Human-Inspired Defense Strategy
Exploiting LLM Security Vulnerabilities in Structured Outputs
How prefix-tree mechanisms can be manipulated to bypass safety filters
Defending LLMs Against Jailbreaking
Efficient safety retrofitting using Direct Preference Optimization
The Anchored Safety Problem in LLMs
Why safety mechanisms fail in the template region
Breaking the Jailbreakers
Understanding defense mechanisms against harmful AI prompts
Bypassing LLM Safety Measures
New attention manipulation technique creates effective jailbreak attacks
Fortifying LLM Defenses
Systematic evaluation of guardrails against prompt attacks
Real-Time Jailbreak Detection for LLMs
Preventing harmful outputs with single-pass efficiency
Defending LLMs Against Jailbreak Attacks
A Novel Safety-Aware Representation Intervention Approach
Exposing the Mousetrap: Security Risks in AI Reasoning
How advanced reasoning models can be more vulnerable to targeted attacks
Evolving Prompts to Uncover LLM Vulnerabilities
An automated framework for scalable red teaming of language models
The Safety Paradox in LLMs
When Models Recognize Danger But Respond Anyway
Exploiting LLM Vulnerabilities
How psychological priming techniques can bypass AI safety measures
Improving Jailbreak Detection in LLMs
A guideline-based framework for evaluating AI security vulnerabilities
Breaking the Guardrails: LLM Security Testing
How TurboFuzzLLM efficiently discovers vulnerabilities in AI safety systems
Essence-Driven Defense Against LLM Jailbreaks
Moving beyond surface patterns to protect AI systems
The FITD Jailbreak Attack
How psychological principles enable new LLM security vulnerabilities
The Vulnerable Depths of AI Models
How lower layers in LLMs create security vulnerabilities
Flowchart-Based Security Exploit in Vision-Language Models
Novel attack vectors bypass safety guardrails in leading LVLMs
Bypassing LLM Safety Guardrails
How Adversarial Metaphors Create New Security Vulnerabilities
Defending LLMs Against Multi-turn Attacks
A Control Theory Approach to LLM Security
The Safety-Reasoning Tradeoff
How Safety Alignment Impacts Large Reasoning Models
Breaking the Safety Guardrails
How Language Models Can Bypass Security in Text-to-Image Systems
Cross-Model Jailbreak Attacks
Improving attack transferability through constraint removal
The Hidden Power of 'Gibberish'
Why LLMs understanding unnatural language is a feature, not a bug
Detecting LLM Jailbreaks through Geometry
A novel defense framework against adversarial prompts
Fortifying LLM Defenses
A dual-objective approach to prevent jailbreak attacks
Fortifying Multimodal LLMs Against Attacks
First-of-its-kind adversarial training to defend against jailbreak attempts
From Complex to One-Shot: Streamlining LLM Attacks
Converting multi-turn jailbreak attacks into efficient single prompts
Beyond Refusal: A Smarter Approach to LLM Safety
Teaching AI to explain safety decisions, not just say no
Securing Multimodal AI Systems
A Probabilistic Approach to Detecting and Preventing Jailbreak Attacks
Exploiting Dialogue History for LLM Attacks
How attackers can manipulate conversation context to bypass safety measures
Strengthening LLM Safety Through Backtracking
A novel safety mechanism that intercepts harmful content after generation begins
Jailbreaking LLMs Through Fuzzing
A new efficient approach to detecting AI security vulnerabilities
Breaking Through LLM Defenses
How Tree Search Creates Sophisticated Multi-Turn Attacks
Certified Defense for Vision-Language Models
A new framework to protect AI models from visual jailbreak attacks
Progressive Defense Against Jailbreak Attacks
A novel approach to dynamically detoxify LLM responses
Defending AI Vision Systems Against Attacks
A tit-for-tat approach to protect multimodal AI from visual jailbreak attempts
Breaking Through AI Guardrails
New Method Efficiently Bypasses LVLM Safety Mechanisms
MirrorGuard: Adaptive Defense Against LLM Jailbreaks
Using entropy-guided mirror prompts to protect language models
Sentinel Shield for LLM Security
Real-time jailbreak detection with a single-token approach
Bypassing LLM Code Safety Guardrails
How implicit malicious prompts can trick AI code generators
Systematic Jailbreaking of LLMs
How iterative prompting can bypass AI safety guardrails
The JOOD Attack: Fooling AI Guardrails
How Out-of-Distribution Strategies Can Bypass LLM Security
Bypassing LLM Safety Guardrails
How structured output constraints can be weaponized as attack vectors
The Security-Usability Dilemma in LLM Guardrails
Evaluating the tradeoffs between protection and functionality
Breaking Through Linguistic Barriers
How Multilingual and Accent Variations Compromise Audio LLM Security
PiCo: Breaking Through MLLM Security Barriers
A progressive approach to bypassing defenses in multimodal AI systems
Defending LLMs Against Jailbreak Attacks
A Lightweight Token Distribution Approach for Enhanced Security
Securing LLMs Against Harmful Outputs
A Novel Representation Bending Approach for Enhanced Safety
Securing the Guardians: The LLM Security Evolution
Analyzing jailbreak vulnerabilities and defense strategies in LLMs
Mapping LLM Vulnerabilities
A Systematic Classification of Jailbreak Attack Vectors
The Sugar-Coated Poison Attack
How benign content can unlock dangerous LLM behaviors
Humor as a Security Threat
How jokes can bypass LLM safety guardrails
Strengthening LLM Safety Guardrails
A reasoning-based approach to balance safety and usability
AdaSteer: Adaptive Defense Against LLM Jailbreaks
Dynamic activation steering for stronger LLM security with fewer false positives
Securing Reasoning Models Without Sacrificing Intelligence
Balancing safety and reasoning capability in DeepSeek-R1
The Cost of Bypassing AI Guardrails
Measuring the 'Jailbreak Tax' on Large Language Models
Vulnerabilities in LLM Security Guardrails
New techniques bypass leading prompt injection protections
