Free Chinese AI models nearly match OpenAI's best at hacking networks (opens the source in a new tab)
- Vals, which measures AI capabilities, set models loose on mock corporate networks. OpenAI's gated cyber model Daybreak completed 81% of tasks; the open-weight DeepSeek managed 73% and tied it on some.
- Vals says open models do the job for about a tenth of the cost, under $3 a run, partly because they can run on one's own hardware.
- That undercuts the idea, behind Anthropic's limited release of Claude Mythos Preview, that gated access buys defenders much time. Real attackers still mostly use older, cheaper models.
- 73%
- DeepSeek's share of hacking tasks completed
Why it matters Defenders can no longer assume top-tier hacking help sits behind trusted-access programs. Patching and network monitoring now have to plan for anyone with a downloaded model having near-frontier tools.
Other side Mock networks are not real breaches, and an Edera co-founder notes the closed labs probably hold stronger unreleased models. Cheap open models also help defenders, who can run them to find holes first.