What happened
Researchers at the AI evaluation firm Vals conducted simulated network security tests comparing the performance of closed-source AI models against open- alternatives. In scenarios involving complex, multi-step network instructions—such as compromising a corporate network via a public-facing website—OpenAI’s proprietary 'Daybreak' model achieved an 81% success rate. The open-weight DeepSeek model followed closely with a 73% success rate, with the two models performing identically on several specific tasks. The study also highlighted a significant cost disparity: executing these tasks with the proprietary GPT-6-Sol model costs between $17 and $46 in tokens, whereas the open-weight DeepSeek model costs between $2 and $3, as it can be deployed on local hardware.
The evaluation, conducted by Vals, tested AI models on their ability to navigate and compromise simulated corporate network environments. The tests included scenarios where models had to traverse between two distinct company networks.
OpenAI's Daybreak model, currently restricted to trusted partners, served as the for high-end performance, completing 81% of assigned tasks. The open- DeepSeek model achieved a 73% success rate, demonstrating that the performance gap between proprietary and open-weight systems is narrowing significantly.
Beyond performance, the study highlighted a stark economic difference. Running these tasks on proprietary models like GPT-6-Sol incurs costs of $17–$46 per session, while the open- DeepSeek model costs only $2–$3, largely due to the ability to run the model on local infrastructure.
The findings suggest that the strategy of restricting access to frontier models to prevent the spread of advanced hacking capabilities is becoming less effective as open- alternatives reach similar levels of proficiency.
Why it matters
The findings challenge the prevailing industry strategy of 'gated access' for frontier AI models, which was intended to delay the proliferation of advanced cyber-offensive capabilities. By demonstrating that open- models can achieve near-parity with restricted, closed-source systems, the research suggests that the window for controlling the spread of these capabilities is closing faster than anticipated. This shift forces a re-evaluation of security postures for both defenders and potential targets. While the cost-efficiency of open-weight models provides a clear incentive for security professionals to adopt them for defensive testing, it simultaneously lowers the barrier to entry for malicious actors. The results indicate that the 'head start' strategy pursued by labs like Anthropic—which restricted access to models like Claude Mythos Preview—may be losing its efficacy as high-performance, open-weight alternatives become more accessible. Ultimately, this development underscores a transition toward a landscape where advanced AI-driven hacking tools are becoming a commodity, necessitating a shift in how organizations approach vulnerability management and network defense.
The research validates concerns that the 'gated access' model—a cornerstone of policy for frontier labs—is being bypassed by the rapid advancement of open- models.
For defenders, the lower cost and high performance of open- models offer a powerful tool for red-teaming and proactive security testing. However, this same accessibility provides a significant advantage to attackers who can now deploy sophisticated capabilities without the oversight or costs associated with proprietary APIs.
The study confirms that the transition to a world where advanced cyber-offensive AI is broadly available has arrived on the timeline predicted by researchers earlier this year.
The debate continues regarding whether these tools represent a new category of threat or an acceleration of existing ones, but the consensus is that the speed and scale of vulnerability discovery are increasing, putting pressure on organizations to patch systems faster.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
The primary concern for the coming months is whether the accessibility of these high-performance open- models will lead to a measurable increase in sophisticated, AI-automated cyberattacks. While current observations of state-backed and criminal hacking groups show a tendency to conserve resources by using older, less capable models, the availability of cheaper, more powerful tools may change this behavior. Security experts will be monitoring whether the automation of vulnerability discovery and network probing leads to a fundamental shift in the threat landscape or merely amplifies existing security challenges. Additionally, the industry will likely watch for how AI labs adjust their release strategies in response to the diminishing effectiveness of model gating.
Watch for whether criminal or state-sponsored actors shift from using older, legacy models to these more capable open- alternatives for their operations.
Monitor how AI labs respond to these findings; if gating is no longer an effective deterrent, labs may pivot toward different defensive strategies or increased investment in AI-native security monitoring.
Observe the impact on corporate security budgets and practices as the cost of running advanced AI-driven security simulations drops, potentially leading to more frequent and rigorous internal testing.
Track the ongoing debate among cybersecurity experts regarding whether AI has fundamentally altered the principles of defense or if it remains a tool that requires human oversight to be truly effective.