ValsAI 网络攻防测试：V4.1 Flash 追平 GPT-6 Sol

Mirolim Mirzakhmedov @MrMirolim原推发布于 2026.10.21

在 X 上打开原推

ValsAI tested the stock DeepSeek V4.1 Flash 1017 on their simulated corporate network cyber benchmark, and it matched GPT-6 Sol for a fraction of the cost. Now imagine what happens when someone runs that exact task using the abliterated community quants of 1017 where the refusal circuits are completely stripped. People on here love to chant "we don't need regulation, just give us open weights" (while conveniently confusing open weights with open source). They celebrate the download link without imagining the downstream consequences. When an uncensored, dirt-cheap model can autonomously compromise a corporate network, "free weights" stops being an ideology and becomes a systemic vulnerability. There’s a reason top Chinese labs are pulling back from open-weight drops on their most capable models, and Beijing is moving to restrict frontier weight exports. You can rate-limit an API, you can't recall a weight file. We need real safeguards before uncensored cyber agents become plug-and-play, because once one is loose with no guardrails, nobody controls how far it goes, not even the person who launched it. https://x.com/ValsAI/status/2112471082397778083

这条是引用着另一条写的：x.com/valsai/status/2112471082397778083