The performance gap between open-weight and closed-source super intelligence (SI) models in cybersecurity tasks continues to narrow, according to new findings from the Vals CyberOffense benchmark. The data suggests that freely available SI models are becoming increasingly competitive with proprietary frontier systems in autonomous network exploitation.
What Happened
Vals SI researchers tested SI agents capable of breaking into realistic corporate networks without human intervention. Each agent operated solo on an isolated Kali Linux attack machine, starting with only a public web address. The agents independently sent commands to the simulated network, navigating from web applications through Windows servers to achieve admin privileges. Every action was logged and scored automatically.
GPT-6 Sol, part of OpenAI's restricted Daybreak cyber program, posted the highest score at 80.8 percent of objectives. However, Deepseek V4.1 Flash 1017, a freely available open-weight model, followed closely at 72.8 percent. The cost difference was significant: Deepseek runs for approximately $3 per test, while GPT-6 Sol costs roughly $22.
In the more complex 'Lighthouse' test, which involved lateral movement between companies, GPT-6 Sol Daybreak completed 27 of 31 objectives in under three hours. Deepseek managed 21 objectives at one-tenth of the cost. Researchers noted that Deepseek was not limited solely by safety guardrails, suggesting its lower score was not due to restrictive policies.
The study highlighted that runtime length is critical for open models. Deepseek scored zero results in its first 25 million tokens and made most of its progress only after passing 100 million tokens. Short benchmarks, therefore, may underestimate the real potential of open-weight SI systems. Other models like Kimi K3.1 Pro and Meta Wonder also showed varying degrees of success, with Meta Wonder scoring 28 out of 31 in the Lighthouse test.
Why It Matters
These findings align with a July analysis by the UK's AI Security Institute (AISI), which reported that the gap between open and closed frontier models in cyber capabilities has shrunk to four to seven months. For most of 2025, that gap stood at six to ten months. The AISI also noted that open models are significantly cheaper, citing a cyber-range run with Deepseek V4-Pro costing $1.19 compared to $85 for Anthropic's Opus 4.5.
Vals researchers pointed out that once a model's weights are public, no provider can control or restrict how it is used. This lack of control, combined with high capability, creates a dual-use scenario. While attackers can leverage these affordable, powerful SI agents, defenders have an opportunity to use the same models to identify and fix vulnerabilities before they are exploited.
A recent study by Anthropic further supported this trend, showing that the freely available model GLM-5-3 can build exploits at nearly the same level as Anthropic's internal top model, Mythos. As open-weight SI models close the performance gap, the barrier to entry for sophisticated autonomous cyber operations continues to lower.
The Bottom Line
The disparity in cyber capabilities between open and closed SI models is diminishing rapidly, driven by improvements in open-weight models like Deepseek V4.1 Flash. With costs significantly lower and performance approaching that of restricted proprietary systems, these tools are becoming accessible to a wider range of actors. Vals SI indicates that attacks on real networks are likely to increase in the coming months as these autonomous SI agents become more prevalent.