The field of large language models (LLMs) is rapidly evolving, with a growing focus on safety and security. Recent developments have highlighted the vulnerability of LLMs to various types of attacks, including jailbreaks, prompt injection, and data poisoning. To address these concerns, researchers have proposed innovative defense mechanisms, such as contextual integrity verification, latent fusion jailbreak detection, and safe-completion training. These approaches aim to improve the robustness of LLMs against adversarial attacks while maintaining their helpfulness and performance. Noteworthy papers in this area include those that propose novel evaluation frameworks for jailbreak attacks, develop more effective attack methods, and investigate the role of context filtering in maintaining safe alignment of LLMs. Overall, the field is moving towards a more comprehensive understanding of LLM safety and security, with a focus on developing practical and effective solutions to mitigate potential risks.
Advances in Large Language Model Safety and Security
Sources
Universally Unfiltered and Unseen:Input-Agnostic Multimodal Jailbreaks against Text-to-Image Model Safeguards
DINA: A Dual Defense Framework Against Internal Noise and External Attacks in Natural Language Processing
Guardians and Offenders: A Survey on Harmful Content Generation and Safety Mitigation
SLIP: Soft Label Mechanism and Key-Extraction-Guided CoT-based Defense Against Instruction Backdoor in APIs
Beyond Uniform Criteria: Scenario-Adaptive Multi-Dimensional Jailbreak Evaluation
LLM Robustness Leaderboard v1 --Technical report
Quantifying Conversation Drift in MCP via Latent Polytope
Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMs
LLM Unlearning Without an Expert Curated Dataset
Many-Turn Jailbreaking
Towards Effective Prompt Stealing Attack against Text-to-Image Diffusion Models
Who's the Evil Twin? Differential Auditing for Undesired Behavior
A Real-Time, Self-Tuning Moderator Framework for Adversarial Prompt Detection
Multi-Turn Jailbreaks Are Simpler Than They Seem
Securing Educational LLMs: A Generalised Taxonomy of Attacks on LLMs and DREAD Risk Assessment
Towards Effective MLLM Jailbreaking Through Balanced On-Topicness and OOD-Intensity
From Hard Refusals to Safe-Completions: Toward Output-Centric Safety Training
Can AI Keep a Secret? Contextual Integrity Verification: A Provable Security Architecture for LLMs
An Audit and Analysis of LLM-Assisted Health Misinformation Jailbreaks Against LLMs
Latent Fusion Jailbreak: Blending Harmful and Harmless Representations to Elicit Unsafe LLM Outputs
Context Misleads LLMs: The Role of Context Filtering in Maintaining Safe Alignment of LLMs
The Cost of Thinking: Increased Jailbreak Risk in Large Language Models
Jailbreaking Commercial Black-Box LLMs with Explicitly Harmful Prompts