A recent study by Vals, a firm that benchmarks artificial-intelligence capabilities, indicates that publicly available AI systems are now almost as proficient at compromising network environments as the most advanced proprietary models. The analysis marks a turning point for defenders who have relied on the presumed superiority of closed-source tools to delay widespread misuse.
The researchers constructed two synthetic environments to gauge performance. In the first, a mock corporate network was reachable through a public website, requiring the model to navigate typical external attack vectors. The second scenario linked two separate company networks, forcing the AI to first breach one system before leveraging that foothold to infiltrate the other, thereby testing multi-stage intrusion tactics.
Among the models examined, OpenAI’s Daybreak,restricted to trusted partners,achieved the highest success rate, completing 81 percent of assigned objectives on average. The Chinese open-weight model DeepSeek followed closely, reaching 73 percent, and on several tasks the two systems produced identical outcomes, underscoring the narrowing gap between open and closed solutions.
Cost efficiency emerged as a decisive factor. Vals reported that executing an attack with the proprietary GPT-6-Sol model required between $17 and $46 in token expenditure, whereas DeepSeek performed comparable actions for merely $2 to $3. Because open models can run on locally owned hardware, they avoid the recurring fees associated with cloud-based services, a point highlighted by Vals CEO Rayan Krishnan as a strong incentive for defensive teams.
The broader cybersecurity landscape has already seen AI accelerate routine hacking steps, enabling both attackers and defenders to discover large numbers of software flaws and probe networks at unprecedented speed. Experts continue to debate whether this automation reshapes the fundamental principles of security or simply amplifies challenges that have existed for decades.
The findings cast doubt on the effectiveness of gating access to cutting-edge models. As Nikil Ravi of Vals noted, “The best frontier models… are now starting to be matched by open models that anybody can download.” Earlier statements from a U.S. lab researcher to WIRED warned that capabilities could become broadly available within six to twenty-four months, suggesting that current restriction strategies may only buy limited time.
Observations of actual cyber-crime activity reveal that many threat actors still favor older, less capable models to conserve resources, even as they occasionally employ newer tools. Nonetheless, defenders must assume that well-funded adversaries have already accessed some closed models, reinforcing the need for comprehensive defensive postures that do not rely solely on model secrecy.