New research from Vals shows that open-weight AI models from China, particularly DeepSeek and GLM, are significantly narrowing the cyberattack capability gap compared to leading proprietary US AI models.
Vals evaluated the ability to execute multi-step cyberattacks on simulated enterprise systems, including web intrusion and lateral movement across internal networks.
OpenAI’s advanced cybersecurity model Daybreak led with an 81% task completion rate, while DeepSeek achieved 73%. On certain tasks, the two models achieved equivalent results.
The cost disparity is particularly striking. According to Vals, GPT-6-Sol costs $17–$46 per test attack, whereas DeepSeek requires only $2–$3.
Open-weight models can be downloaded and operated on private infrastructure, lowering costs and reducing vendor control over access rights.
These findings pose a challenge to OpenAI’s and Anthropic’s strategies of restricting access to AI models with high cyberattack capabilities.
Anthropic previously predicted that advanced cyberattack capabilities would become widespread within 6 to 24 months.The new research shows this dissemination is happening rapidly.
However, test results do not yet prove equivalent efficacy in real-world environments. Many threat groups continue to use older, low-cost models that adequately meet their objectives.
Experts warn that weight manipulation to disable safety mechanisms could further increase the risk of AI misuse.
📌 Conclusion: Vals’ research demonstrates that advanced cyberattack capabilities are becoming more accessible due to open-weight AI models and low operating costs. The 81% versus 73% gap shows that the advantage of proprietary models is no longer absolute in these benchmarks. This trend necessitates shifting the focus of cybersecurity from controlling model access to proactively discovering vulnerabilities, automating defense, and shortening remediation times before attackers can exploit them.

