General

Open-Weight Chinese AI Models Close the Cybersecurity Gap with Frontier Labs

New research shows open-weight Chinese AI models like DeepSeek are now nearly as effective as closed frontier models in executing complex cyberattacks, significantly lowering the cost and barrier to entry for malicious actors.

S
SciTech Editorial Team
Newsroom
··2 min read
Open-Weight Chinese AI Models Close the Cybersecurity Gap with Frontier Labs
A recent study by Vals reveals that open-weight artificial intelligence models are now nearly as capable of hacking into computer networks as the most advanced closed models, marking a significant turning point in cybersecurity. The research indicates that the definitive edge previously held by proprietary frontier models is eroding, potentially rendering recent efforts by major labs to restrict access to their tools moot. Vals tested various AI models on mock corporate networks, assigning them complex, multi-step network instructions. In the first scenario, models accessed a mock corporate network via a public website; in the second, they had to compromise one network to gain access to another. The results showed that OpenAI’s advanced cyber model, Daybreak, completed 81% of tasks on average. However, the Chinese open-weight model DeepSeek was not far behind, achieving a 73% success rate, with both models scoring equally on several specific tasks. Beyond capability, the study highlights a stark cost disparity. Executing an attack using OpenAI’s GPT-6-Sol costs between $17 and $46 in tokens, whereas DeepSeek costs only $2 to $3. This affordability, combined with the ability to run open models on local hardware, provides a strong incentive for defenders to adopt open-source solutions. Nikil Ravi, a researcher at Vals, noted that frontier models are now being matched by open models that anyone can download, signaling a shift in the balance of power. The findings align with predictions made by Anthropic’s red team lead, Logan Graham, who warned in April that the industry needed to prepare for a world where such capabilities would be broadly available within 6 to 24 months. Seven months later, that transition appears to have arrived on schedule. While experts debate whether AI fundamentally alters cybersecurity tenets or merely amplifies existing challenges, the practical implication is clear: criminal hackers and state-backed actors now have access to powerful, low-cost tools. Alex Zenla, CTO of Edera, agrees that open-weight models are capable of producing advanced attacks today, especially when safety training is disabled. However, he cautions that closed labs may have trained unreleased models with even greater capabilities, describing the current landscape as a "cat-and-mouse game." As open models close the gap, the security landscape becomes increasingly complex, requiring defenders to assume that sophisticated attackers already have access to the most advanced tools.

Share

More from this section: General