The field of large language models is rapidly evolving, with a growing focus on improving safety and security. Recent research has highlighted the vulnerabilities of these models to adversarial attacks, data breaches, and malicious fine-tuning. In response, researchers have developed innovative methods to enhance safety, including reinforcement learning approaches, safety-aware probing optimization, and backdoor detection frameworks. Noteworthy papers in this area include AutoRAN, which demonstrates the effectiveness of weak-to-strong jailbreaking attacks, and CTRAP, which proposes a novel paradigm for inducing model collapse to prevent harmful fine-tuning. Overall, the field is moving towards developing more robust and secure large language models that can mitigate potential risks and ensure safe deployment in real-world applications.
Advances in Safety and Security of Large Language Models
Sources
AutoRAN: Weak-to-Strong Jailbreaking of Large Reasoning Models
GenoArmory: A Unified Evaluation Framework for Adversarial Attacks on Genomic Foundation Models
Noise Injection Systemically Degrades Large Language Model Safety Guardrails
Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activations
Safety Subspaces are Not Distinct: A Fine-Tuning Case Study
Context Reasoner: Incentivizing Reasoning Capability for Contextualized Privacy and Safety Compliance via Reinforcement Learning
SifterNet: A Generalized and Model-Agnostic Trigger Purification Approach
sudoLLM : On Multi-role Alignment of Language Models
Linear Control of Test Awareness Reveals Differential Compliance in Reasoning Models
SAFEPATH: Preventing Harmful Reasoning in Chain-of-Thought via Early Alignment
HAVA: Hybrid Approach to Value-Alignment through Reward Weighing for Reinforcement Learning
How Should We Enhance the Safety of Large Reasoning Models: An Empirical Study
Be Careful When Fine-tuning On Open-Source LLMs: Your Fine-tuning Data Could Be Secretly Stolen!
Hierarchical Safety Realignment: Lightweight Restoration of Safety in Pruned Large Vision-Language Models
SafeKey: Amplifying Aha-Moment Insights for Safety Reasoning
Three Minds, One Legend: Jailbreak Large Reasoning Model with Adaptive Stacked Ciphers
CTRAP: Embedding Collapse Trap to Safeguard Large Language Models from Harmful Fine-Tuning
Finetuning-Activated Backdoors in LLMs
Mitigating Fine-tuning Risks in LLMs via Safety-Aware Probing Optimization
Accidental Misalignment: Fine-Tuning Language Models Induces Unexpected Vulnerability
Backdoor Cleaning without External Guidance in MLLM Fine-tuning