Today, I am leaving Tencent's Xuanwu Lab, which I joined immediately after graduation and where I have worked for seven years. I will soon join MiniMax and begin a new journey. At this important turning point in my life, I want to record the ideas that have guided me through these years and continue to shape my aspirations.

Anthropic launched Project Glasswing and Claude Mythos Preview, demonstrating formidable automated vulnerability capabilities across OpenBSD, FFmpeg, the Linux kernel, and more. This post asks where security practitioners should anchor their value as execution-layer skills become increasingly automatable—from “doing the work” to “deciding what work matters,” why higher-level judgment resists replacement longer, and how mastering AI agents becomes the new universal baseline.