The recent developments in the field of AI safety and robustness have seen a significant shift towards creating more resilient and responsible AI systems, particularly in high-stakes domains such as chemistry and robotics. Researchers are increasingly focusing on benchmarking and evaluating the safety and accuracy of large language models (LLMs) in specialized fields, such as chemistry, to ensure they do not generate harmful or incorrect responses. This trend is exemplified by the introduction of benchmarks like ChemSafetyBench, which rigorously tests LLMs on their ability to handle complex chemical queries safely. Additionally, there is a growing emphasis on integrating safety mechanisms directly into AI systems, particularly in robotic manipulation, where the potential for harm is high. The concept of 'Safety-as-policy' in robotic manipulation, which involves training robots to anticipate and avoid hazards, represents a notable advancement in this direction. Furthermore, the field is exploring the potential of neuroscience to inform AI safety, with a roadmap highlighting how emulating the brain's architecture and learning algorithms could lead to more robust and cooperative AI systems. Overall, the field is moving towards a more holistic approach to AI safety, integrating domain-specific benchmarks, direct safety implementations, and insights from neuroscience to create AI systems that are not only powerful but also safe and responsible.