The field of large language models is moving towards developing more robust and reliable safety mechanisms to mitigate potential risks and harms. Recent research has focused on creating innovative solutions to address the challenges of ensuring safety and usefulness in complex multimodal settings. One of the key directions is the development of streaming risk detection frameworks that can accurately identify and mitigate harmful content in real-time. Another important area of research is the creation of benchmarks and evaluation metrics to assess the safety and robustness of large language models. Noteworthy papers in this area include: Kelp, which proposes a novel plug-in framework for streaming risk detection, and DUAL-Bench, which introduces a multimodal benchmark for measuring over-refusal and robustness in vision-language models. SafeMT is also notable for its introduction of a benchmark for multi-turn safety in multimodal language models. Protect presents a natively multi-modal guardrailing model designed for enterprise-grade deployment, and Risk-adaptive Activation Steering proposes a method for safe multimodal large language models. Qwen3Guard introduces a series of multilingual safety guardrail models with specialized variants for generative and streaming safety classification.