给出 AI Agent 自动化利用的能力边界,评估攻击门槛可直接引用
llmai_agent内核
apart from ip options potentially being dropped in transit, kernel text leaks taking time, and exploit reliability varying across configurations, for everyone to have peace of mine:
open models can't reliably exploit this vulnerability on its own without an experienced operator guiding it.
so you're all fine, php2shell is considerably easier to exploit than this, and even open models struggled with that. it is probably only openai and anthropic internal frontier models that can reliably pull this off autonomously :)