gpt-6POST
Security analysis test: DeepSeek V4.1 Flash edges out GPT-6 Astra
A security researcher compared several models on the same harness and found DeepSeek V4.1 Flash produced better output than GPT-6 Astra and Grok 4.6, though none were perfect.
Cos(余弦)😶🌫️
@evilcosArchived 2026.09.15Originally posted 2026.09.14
Cos(余弦)😶🌫️ @evilcosOriginally posted 2026.09.14
最近几个安全分析，DeepSeek V4.1 Flash 表现的不错，同样本下，居然比 gpt-6-astra/grok-4.6 输出的要好些，比 fable-5.1 差了一点。 当然，都没有完美的分析。 这个粗略的对比仅供参考，估计关键点在我做的 harness 工程上。
x.com/evilcos/status/2099514450483315094Archived 2026.09.15 10:12