Recent evaluations on the Vals AI benchmark demonstrate that leading open-weight and proprietary artificial intelligence models are increasingly close in their offensive cyber capabilities. According to reporting from The Decoder, OpenAI’s GPT-6 Sol Daybreak successfully achieved 80.8 percent of attack objectives within simulated corporate network environments. Meanwhile, DeepSeek V4.1 Flash, an open-weight model, reached 72.8 percent on the same benchmark while operating at a tenth of the cost.
The Shrinking Defender Window
Security analysts note that open-weight models present unique challenges because their underlying weights can be downloaded, modified, and run locally without the restrictions typically imposed by proprietary application programming interfaces. As the performance gap between closed systems and accessible open models narrows significantly, industry observers warn that the operational window for network defenders to adapt their strategies is contracting rapidly.
Unknown Variables And Future Outlook
Despite the benchmark findings, the announcement did not specify the exact parameters of the simulated corporate networks used in the Vals AI test, nor did it detail the specific defensive countermeasures active during the evaluations. It remains unclear how these models perform against real-world enterprise security architectures or what specific mitigation techniques might effectively counter such automated tactics.
As both open and proprietary models continue to advance in offensive utility, security teams must prepare for more accessible and capable automated threats.
Lead image: Brett Sayles / Pexels
