实测：V4.1 Flash 在 Web 漏洞任务上接近饱和，可写内核与 v8 n-day 利用

s1r1us @S1r1u5_原推发布于 2026.10.22

在 X 上打开原推

i want to fearmonger! there's a good chance we'll see a lot more attacks like railway and opencode. i wouldn't even be surprised if attackers are already using deepseek flash. deepseek v4.1 flash is the first open model that made me update insanely hard on the gap between open and frontier models. how the fuck is a 500b flash model this good? it seems like deepseek figured out something similar to the post-training jump from gpt-5.5 to gpt-5.6 that turned sol into a maniac hacker. its small, fast, and reasons insanely well about code. we've been benchmarking it, and it nearly saturates most web vulnerability tasks. it still can't write complex exploits like gpt-5.6 sol, things like kvm escapes, php2shell, etc. but it can write kernel exploits which are not complex, there's a recent n-day that flash one-shots, same with v8 ndays. if deepseek scales whatever they figured out with flash to v4.1 pro, i wouldn't be surprised to see sol- ike results. this thing can be abliterated and run on fairly modest gpu hardware, and they released an inference engine to blast tokens out.