Open-weight models close the cyber gap: DeepSeek V4.1 Flash hits 72.8% at a tenth the cost
The Decoder · rss · 2026-11-07
Per The Decoder, the Vals AI benchmark shows the gap between open and closed models in cyber capabilities shrinking: OpenAI's GPT-6 Sol Daybreak achieves 80.8% of attack objectives in simulated corporate networks, while open-weight DeepSeek V4.1 Flash reaches 72.8% at a tenth of the cost. Analysts warn the window for defenders is closing fast, since open models can't be restricted once released.
More from Models
- GLM and DeepSeek accused of benchmark awareness after MaxxBench rewritten-task failures — teortaxesTex · 2026-11-08
- Kambhampati: Frontier models' 'jagged intelligence' traces back to opaque training data mixes — rao2z · 2026-11-08
- Researcher flags scaling fit that hard-codes capability jump to zero — JacquesThibs · 2026-11-08
- Hermes is still massively underrated, users say it beats Grok on freedom — Teknium · 2026-11-08
- Two Minute Papers Explains the OpenAI Astra Weight-Space Programming Hack — Two Minute Papers · 2026-11-08
- Don't trust LLM 'emergence' claims — read 'emergence' as 'training', researcher tells students — tallinzen · 2026-11-08