Vals says open-weight models nearly match closed models at hacking
Original title: Free Chinese AI Models Are Getting Scary Good at Hacking
On simulated networks, OpenAI’s Daybreak finished 81 percent of tasks and DeepSeek 73 percent, at a much lower token cost.
Vals tested AI models on multi-step tasks against mock corporate networks and found open-weight models approaching the best closed cyber models. OpenAI’s partner-only Daybreak completed 81 percent of tasks on average across two tests; a Chinese open-weight DeepSeek model completed 73 percent, tying Daybreak on some tasks. Vals reported token costs of $17 to $46 for GPT-6-Sol versus $2 to $3 for DeepSeek. The result comes seven months after Anthropic framed gated access, starting with Claude Mythos Preview, as buying defenders time. Observers say real attackers often still use older models, and closed labs may hold stronger unreleased systems.
Unconfirmed: Results are from simulated tasks; whether closed labs hold stronger unreleased models is an expert caveat, not a confirmed finding.
- One scenario used a public website into a mock corporate network.
- The second required compromising one company network to reach another.
- Vals says some tasks were scored equally by Daybreak and DeepSeek.
- OpenAI did not immediately respond to a request for comment.
- Edera’s Alex Zenla also cited abliterated models that disable safety training.
- Countries
- China
From the original article
3 min read at wired.com
An experiment in which AI models were asked to break into simulated networks showed the hacking skills of DeepSeek and GLM to be not far behind OpenAI’s most advanced cyber model, which isn’t public.
By Will Knight, Lily Hay Newman · wired.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.