展示开源模型在真实漏洞挖掘上的能力边界与开销。
llm逆向漏洞挖掘
Open-weights just require some focus. DSV4.1 for RE of regions + K3 || GLM5.3 for deeper contextual RE + VR. Exploit development is doable with open-weights but it does incur a decent amount of overhead. If you aren’t familiar with a specific subsystem or interactions between components then you may have a rough time. With the open weights, I was able to discover 4 within a more hardened surface (oob write, cross-priv info leak, 2x null-deref DoS). With a library of primitives, K3 or GLM5.3 can do well.
Daybreak is excellent though, I would highly recommend it for VR; but I was able to net these 4 reportable bugs in <72 hours with an open-weight setup — so we shouldn’t throw shade at open-weights, they’re pretty great when focused.