你刚刷到这条消息，本来准备顺手划走，但又怕自己错过了真正会影响下一步判断的那一点。
最容易做错的，是NVIDIA / SkillSpector；代价往往是如果只盯表面热闹，你很容易在错误方向上花掉时间、预算和注意力。；我先给一个保守判断：Agent安全先审意图，再审代码。
You see NVIDIA / SkillSpector in your feed, almost scroll past it, then stop because you do not want to miss the one thing that could change your next move. The easy mistake is to treat this like another tool update and keep looking at features. That is how teams waste time, budget, and attention in the wrong place.
My conservative read is simple: audit intent before code.
A release is worth your time only if it changes your next step. This one does. The first check is not “does the code look malicious?
” It is “does the skill’s stated purpose match its permissions and its data flow?” That is the real shift.
The number that flipped this for me: across 67,453 skill samples, semantic agentic risk, meaning a mismatch between stated purpose, permissions, and data flow, showed up in 75.3% of suspicious skills, but only 6.8% of malicious ones [S001]. That is the point. The bigger blind spot is not classic malware.
真正该讨论的是：NVIDIA / SkillSpector