给出模型自主做利用链的能力上限,可评估自动化攻防风险。
llm漏洞挖掘议题
To be clear, I did not say with no tools.
What I am saying is *with tools* -- e.g. with the headless API for the best commercial decompilers, and emulation/virtualization solutions (like Corellium Falcon), etc -- which in total, cost hundreds of thousands of dollars, and with a large sample size of working production grade exploits to take inspiration from, the best open source models are *still* completely incapable of autonomously designing and building production grade exploits/exploit chains against the major mobile attack surfaces. With the very close and careful guidance of an expert exploit developer with years of experience, they can be halfway useful.
They can find bugs, sure. But that is infinitely easier than the work involved with curating, exploiting for full effect (i.e. RCE, SBX Escape or LPE) and chaining.
However, the very best closed models, in their various cyber-vetted versions, are capable of doing so. This only became the case in the last 6 weeks or so.