According to new research from AI testing firm Vals, open-weight artificial intelligence models have substantially closed the performance gap with proprietary frontier models in complex network hacking capabilities. In automated penetration tests on mock corporate networks, Chinese open-weight model DeepSeek completed 73 percent of assigned hacking tasks, approaching the 81 percent success rate of OpenAI’s restricted cyber model, Daybreak.
The study highlights significant economic disparities in executing automated offensive security tasks. Vals reported that running an attack using OpenAI’s GPT-6-Sol costs between $17 and $46 in API tokens, compared to just $2 to $3 in compute costs for DeepSeek running on self-hosted hardware. Researchers noted that the affordability of open-weight models provides a strong economic incentive for defensive cyber security teams.
These findings suggest that defensive timeframes secured by gated API releases—such as Anthropic’s Project Glasswing initiative—are shrinking rapidly. Security experts note that as open models democratize high-level cyber skills, defenders must adapt to a landscape where automated vulnerability probing is widely accessible.
Why it matters
CISOs and security teams must assume advanced AI probing tools are now cheaply available to malicious actors.
Open-weight models offer security operators cost-effective alternatives for automated network defense and penetration testing.
Model developers face growing evidence that access-gating strategies provide only short-term delays against capability proliferation.
Source: wired.com



