Open-Weight AI Parity Erodes Proprietary Security Moats
AI-generated image · US National Wire
New research indicates Chinese open-source models are nearly matching the cyber-offensive capabilities of closed frontier systems, shifting the risk calculus for enterprise security.
From a deals and valuation perspective, the perceived security moat provided by closed-source AI models is rapidly evaporating. New research from Vals, as Wired first reported, indicates that open-weight models—specifically those originating from China—are now nearly as capable of executing complex network intrusions as the most advanced proprietary systems.
According to reporting from Wired, Vals conducted experiments using mock corporate networks to test AI models on multi-step hacking tasks. The results showed that OpenAI’s most advanced cyber model, Daybreak—which is restricted to trusted partners—completed 81 percent of tasks. However, the Chinese open-weight model DeepSeek followed closely, completing 73 percent of objectives. Vals researchers noted that the two models performed equally on certain tasks.
This convergence suggests that the strategy of gated access, employed by U.S. labs to slow the proliferation of hacking capabilities, may be losing its efficacy. Wired notes that Anthropic previously premised the limited release of models like Claude Mythos Preview on the belief that gated access would buy time for defenders. Logan Graham, Anthropic's frontier red team lead, told Wired in April that Project Glasswing was intended to give defenders a head start before such capabilities became broadly available within 6 to 24 months.
Beyond raw capability, the economic profile of these tools creates a significant shift in operational costs. Vals reports that using GPT-6-Sol for an attack costs between $17 and $46 in tokens, whereas DeepSeek costs only $2 to $3. Rayan Krishnan, CEO of Vals, suggests this price disparity provides a strong incentive for defenders to pivot toward open models.
Despite these advancements, the real-world application of these tools remains nuanced. Wired reports that cybercriminals and state-backed actors have continued to rely on older, less capable models for most activities. However, Alex Zenla, cofounder and CTO of Edera, warned Wired that the modification of weights to disable safety training in open-weight models could signal the start of a more dangerous security landscape, describing the current environment as a "cat-and-mouse game" between closed labs and open-source developments.