Open-weight artificial intelligence models have reached a level of capability that closely rivals the most advanced closed systems in cybersecurity tasks, according to new research. This development marks a significant shift in the digital security landscape, suggesting that the gap between freely available tools and restricted, high-end models is narrowing rapidly.
Vals, a company that measures AI capabilities, conducted tests to evaluate how well different AI models could execute complex, multi-step network instructions on simulated corporate environments. The findings indicate that closed models no longer hold a definitive advantage in cyber skills. This outcome challenges recent efforts by major technology labs to limit access to their most powerful models, as these restrictions may no longer effectively slow the spread of advanced hacking capabilities.
Performance of Open AI Models in Simulated Attacks
The research tested AI models in two distinct simulated scenarios. The first involved a mock corporate network accessible through a public-facing website. The second scenario required a model to compromise one company network to gain access to a second, separate network. These tests were designed to mimic real-world conditions where attackers move laterally through systems.
OpenAI’s most advanced cyber model, Daybreak, which is available only to trusted partners, performed the best. It completed 81 percent of the assigned tasks in an average of two tests. The Chinese open-weight model DeepSeek was not far behind, completing 73 percent of the objectives on average. Vals researchers noted that on several specific tasks, the two models scored equally.
Nikil Ravi, a researcher at Vals, stated that the best frontier models, which are locked inside labs and available only through trusted access programs, are now starting to be matched by open models that anyone can download. This parity suggests that the transition to a world where advanced hacking capabilities are broadly available has arrived on schedule.
Cost and Accessibility of AI Cyber Tools
One of the most notable findings from the study is the significant cost difference between using open-weight models and proprietary systems for cyber operations. Vals reported that it costs between $17 and $46 in tokens for GPT-6-Sol to carry out an attack. In contrast, the same task using DeepSeek costs only $2 to $3.
Open models are generally cheaper because they can be run on a user’s own hardware rather than relying on expensive cloud APIs. Rayan Krishnan, CEO of Vals, said this cost advantage provides a strong incentive for those playing defense to use open models rather than proprietary ones. This economic factor could accelerate the adoption of these tools across both offensive and defensive cybersecurity teams.
The results align with predictions made earlier in the year by Anthropic, a leading AI lab. Logan Graham, Anthropic’s frontier red team lead, told WIRED in April that the company’s Project Glasswing was focused on giving trusted defenders a head start. He warned that the industry needed to prepare for a world where these capabilities are broadly available within six to 24 months. Seven months later, that timeline appears to have been met.
While AI is automating core hacking tasks and helping both attackers and defenders find vulnerabilities at scale, experts are still debating whether AI has fundamentally altered the conceptual foundation of cybersecurity defense. Real-world observations show that attackers often conserve resources by using older, less capable models for most activities. However, the availability of high-performing open models means that even well-funded defenders must assume that motivated attackers have access to these powerful tools.
Alex Zenla, cofounder and chief technology officer of the cloud security firm Edera, agreed that open-weight models are capable of producing advanced attacks today. He noted that this is especially true for models that have been modified to disable safety training. Zenla described the situation as a cat-and-mouse game, where labs try to outpace open-weight models to secure vulnerabilities before they are widely exploited.
Source: Wired
