Recent research indicates that open-weight artificial intelligence models have reached a level of performance that rivals the most advanced, restricted models in the domain of cybersecurity. A study conducted by the firm Vals evaluated how effectively various AI systems could execute complex, multi-step instructions within simulated corporate network environments. The findings suggest that the gap between proprietary, gated models and publicly available, open-weight alternatives is closing rapidly, potentially undermining strategies that rely on limiting access to advanced AI to prevent the proliferation of cyber-offensive capabilities.
The evaluation involved two distinct scenarios. In the first, models were tasked with infiltrating a mock corporate network via a public-facing website. The second scenario required models to compromise an initial network to gain unauthorized access to a secondary, interconnected system. Researchers measured performance based on the success rate of completing these multi-step objectives.
According to the data, OpenAI’s restricted model, Daybreak, which is currently limited to trusted partners, demonstrated the highest capability, successfully completing 81 percent of assigned tasks. However, the open-weight model DeepSeek followed closely, achieving a 73 percent success rate. In several specific tasks, the performance of the two models was indistinguishable. Other models, including GLM, also showed significant proficiency in these simulated hacking exercises.
Beyond raw performance, the study highlighted a substantial disparity in operational costs. Running an attack using proprietary models like GPT-6-Sol was estimated to cost between $17 and $46 in tokens. In contrast, utilizing open-weight models such as DeepSeek cost between $2 and $3. Because open-weight models can be deployed on private hardware, they offer a more cost-effective and accessible option for those seeking to automate cyber-offensive tasks.
The rapid advancement of these models marks a milestone that some industry observers had anticipated would occur within a 6-to-24-month window. Previously, frontier AI labs justified the restricted release of their most powerful models by arguing that gated access would provide defenders with a necessary head start to prepare for the eventual democratization of these capabilities. The current findings suggest that this transition to broad availability is now underway.
Despite the technical capability of these models to identify vulnerabilities and probe networks, the broader impact on the cybersecurity landscape remains a subject of debate. While AI is clearly automating core hacking tasks and increasing the speed at which software flaws are discovered and exploited, it is not yet clear if this represents a fundamental shift in the nature of cyber threats or merely an amplification of existing security challenges.
Real-world observations of malicious activity suggest that, for the time being, many attackers continue to rely on older, less sophisticated models to conserve resources. Furthermore, security professionals note that highly motivated and well-funded threat actors likely already possess some level of access to advanced, proprietary tools, regardless of current release restrictions.
The emergence of these high-performing open models complicates the defensive strategy of frontier labs. The industry is currently engaged in a cycle where labs attempt to outpace open-weight developments to identify and patch vulnerabilities before they can be exploited by widely available AI. However, the ability to modify the weights of open models to bypass safety training—a process known as obliteration—further complicates the security outlook, potentially accelerating the development of advanced automated attack vectors.
