jailbreak attacks
Security vulnerabilities in AI models where adversarial inputs cause the model to produce unintended or harmful outputs. Understanding and mitigating these attacks is crucial for ensuring the safe deployment of AI systems.
- ALMGuard: Safety Shortcuts and Where to Find Them as Guardrails for Audio–Language Models
- Adjacent Words, Divergent Intents: Jailbreaking Large Language Models via Task Concurrency
- AdvPrefix: An Objective for Nuanced LLM Jailbreaks
- Analogy-based Multi-Turn Jailbreak against Large Language Models
- Attack via Overfitting: 10-shot Benign Fine-tuning to Jailbreak LLMs
- Bits Leaked per Query: Information-Theoretic Bounds for Adversarial Attacks on LLMs
- CoP: Agentic Red-teaming for Large Language Models using Composition of Principles
- GASP: Efficient Black-Box Generation of Adversarial Suffixes for Jailbreaking LLMs
- JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models
- RSafe: Incentivizing proactive reasoning to build robust and adaptive LLM safeguards
- SAFEPATH: Preventing Harmful Reasoning in Chain-of-Thought via Early Alignment
- Short-length Adversarial Training Helps LLMs Defend Long-length Jailbreak Attacks: Theoretical and Empirical Evidence
- T2V-OptJail: Discrete Prompt Optimization for Text-to-Video Jailbreak Attacks
- The VLLM Safety Paradox: Dual Ease in Jailbreak Attack and Defense
- Understanding and Rectifying Safety Perception Distortion in VLMs