Must read
Free Chinese AI Models Are Getting Scary Good at Hacking
The rapid parity of freely downloadable open-weight models in offensive cyber capabilities undermines the effectiveness of gated-access strategies used by frontier labs to prevent misuse.
Reported by Wired ·
What happened
New research from AI evaluation firm Vals shows that the Chinese open-weight model DeepSeek now closely matches restricted frontier models in executing complex, multi-step network hacking tasks. In simulated corporate network attacks, DeepSeek completed 73 percent of objectives at a fraction of the cost compared to 81 percent for OpenAI's gated Daybreak model.
Who it’s about
- Vals
- DeepSeek
- OpenAI
- Anthropic
- Daybreak