GoKawiilOpen-weight AI models now rival closed systems at network hacking, Vals research finds
AI benchmarking firm Vals tested multiple AI models' ability to carry out multi-step intrusions on simulated corporate networks and found that freely downloadable open-weight models now perform nearly as well as top closed models from major labs. Researcher Nikil Ravi said frontier models once restricted to trusted-access programs are being matched by models anyone can download.
GoKawiil's interpretation of the reporting above, not reported fact.
The findings suggest that gating access to advanced AI models, a strategy Anthropic adopted with programs like Project Glasswing to give defenders a head start, may no longer effectively delay the spread of dangerous hacking capabilities. Anthropic's Logan Graham had predicted in April that such capabilities would become broadly available within 6 to 24 months, and Vals' results indicate that timeline may be playing out faster than expected. This could pressure cybersecurity teams and policymakers to rethink defense strategies built around the assumption that only well-resourced labs control the most capable hacking tools.
- Vals' testing found open-weight AI models nearly matching closed frontier models in network-hacking tasks.
- Anthropic had restricted early access to models like Claude Mythos Preview partly to delay proliferation of hacking capabilities.
- The research suggests that timeline for broad availability of advanced hacking AI may be arriving faster than labs anticipated.
Source: wired.com — Will Knight, 2026-11-06
Published there as: “Free Chinese AI Models Are Getting Scary Good at Hacking”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.