DeepSeek Ties OpenAI's Daybreak on Some Cyber Tasks — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Vals tested AI models on two mock corporate-network scenarios, including one that required compromising a first company network before gaining access to a second.
- Daybreak, OpenAI's trusted-partner cyber model, completed an average of 81% of its assigned tasks across two tests, the best result in Vals' benchmark.
- DeepSeek, the Chinese open-weight model, averaged 73% of objectives, tied Daybreak on some tasks and cost $2 to $3 in tokens per attack versus $17 to $46 for GPT-6-Sol.
- Anthropic began limiting new-model access with Claude Mythos Preview in April and gave defenders trusted access through Project Glasswing; frontier red team lead Logan Graham said teams should prepare for broad availability in 6, 12, or 24 months.
- Vals' benchmark findings came alongside a real-world caveat: attackers still conserve resources and mostly rely on older, less capable models, while defenders must assume well-funded, highly motivated attackers already had some access to closed models.
- Alex Zenla said open-weight models already produce advanced attacks and vulnerabilities, particularly abliterated versions with safety training disabled, but cautioned that closed labs have likely trained unreleased models that exceed them.
Why it matters: Daybreak led at 81%, while DeepSeek scored 73%; an attack cost $2 to $3 in tokens with DeepSeek versus $17 to $46 with GPT-6-Sol. That erodes Anthropic's April rationale for buying time through gated releases, while defenders gain a cheaper way to test networks with tools attackers can download just as easily.
Ask SkimNews
.png)
