The safety of large language models relies heavily on the agent harness, which manages various components such as context, memory, and permissions. However, current safety mechanisms often view the harness as a static entity, hindering its ability to adapt to new risks. Researchers have introduced the concept of Trajectory-driven Safety Harness Evolution, which enables the harness to evolve in response to emerging threats1. This approach acknowledges that the harness is not a fixed artifact, but rather a dynamic component that requires continuous updates to ensure the safety of LLM agents. By recognizing the interdependencies between harness components, this method can help identify and mitigate potential safety risks. The evolution of the harness is critical to preventing potential vulnerabilities, so practitioners must prioritize the development of adaptive safety mechanisms to protect against emerging threats.