The field of language model safety is moving towards more dynamic and flexible monitoring approaches. Researchers are exploring ways to mitigate information leakage and improve the evaluation of safety monitors, which is crucial for detecting harmful behaviors in large language models. One of the key directions is the development of adaptive and interpretable monitoring methods that can provide stronger guardrails without wasting resources on easy inputs. Another important area of research is the study of secret knowledge elicitation, which aims to discover knowledge that language models possess but do not explicitly verbalize. Furthermore, there is a growing concern about the safety alignment issues in low-resource language settings, where models can be induced to generate harmful or culturally insensitive content. Noteworthy papers in this area include: Eliciting Secret Knowledge from Language Models, which proposes effective techniques for discovering secret knowledge in language models. Beyond Linear Probes: Dynamic Safety Monitoring for Language Models, which introduces a novel approach to dynamic activation monitoring using Truncated Polynomial Classifiers. OpenAI's GPT-OSS-20B Model and Safety Alignment Issues in a Low-Resource Language, which highlights significant safety concerns in low-resource language settings.