Cybersecurity Gap Between Open and Closed AI Models Narrows

The gap between open and closed AI models in terms of risk when artificial intelligence is exploited for cyberattacks has narrowed to an unprecedented degree. According to an evaluation conducted by AI benchmark assessment organization Vals AI, OpenAI's closed model 'GPT-6 Sol Daybreak' achieved an 80.8% attack success rate in simulated enterprise network environments. Meanwhile, DeepSeek's 'V4.1 Flash', an 'open-weight' model that anyone can access and freely operate, reached a comparable 72.8% success rate, and at just one-tenth the cost of GPT-6.

More important than the numerical figures themselves shown in this evaluation is the significance of the narrowing gap between the two. Closed models can be managed by providing companies in terms of access, which gives them the means to restrict use or adjust features if misuse is detected. However, once open-weight models are released, weight files spread globally, making it nearly impossible for providers to control how they are used. This evaluation precisely demonstrates that such uncontrollable models already possess high cybersecurity attack capabilities in numerical terms.

The Vals AI benchmark targets attack tasks in simulated enterprise network environments. The metric of attack success rate measures how autonomously a model can execute cyberattack procedures, serving as one baseline for assessing how much threat it might pose to real-world security environments. The fact that DeepSeek V4.1 Flash achieves a high 72.8% success rate combined with low cost is noteworthy from the perspective of decreased attack costs.

Concerns about AI-enabled cyberattacks have been raised before, but until now, access to high-performance models has required certain costs and managed procedures, which have functioned as a form of deterrent. As open-weight model performance improves rapidly, such implicit barriers are diminishing. This evaluation can be said to demonstrate that change quantitatively.

Analysts point out that this situation is rapidly narrowing the window of time available for defenders to respond. As AI-powered attacks progress toward automation and lower costs, defenders require resources and expertise to keep pace with these changes. Views are also spreading that the asymmetry between attack and defense may become more pronounced with the proliferation of AI.

A point to watch going forward is how the discussion over how to handle cybersecurity capability evaluations of open-weight models unfolds. Issues such as whether risk assessments should be mandated before model release, or what standards should determine the threshold for danger, are still at a stage where international consensus has not been formed. The existence of quantitative benchmarks such as this one is positioned as material for making that discussion more concrete.

In the field of cybersecurity, AI is also being deployed as a defensive tool. The adoption of AI for threat detection and anomaly analysis is expanding, and the structure of AI being used for both attack and defense is expected to continue. However, as this evaluation demonstrates, the reality that high-performance attack tools are available at relatively low cost within reach brings increased vigilance to defensive efforts.