Chinese artificial intelligence models DeepSeek and Z.ai GLM have almost matched OpenAI in their ability to breach complex corporate networks, a new benchmark study has revealed.
Research conducted by evaluation firm Vals demonstrated that open-weight systems developed in China can execute advanced multi-stage cyberattacks at a level approaching the most capable proprietary technology from the United States.
In the tests, OpenAI's model GPT-6-Sol, an unreleased system developed under the company's closed Daybreak project, achieved the highest overall score by successfully completing 81 per cent of assigned hacking tasks.
China's DeepSeek followed closely with a 73 per cent success rate across the trials, matching the performance of the American model in several individual tests.
Simulated network intrusions
The evaluation subjected the artificial intelligence models to two distinct intrusion scenarios that mirror real cyberattacks. The first challenge required each system to penetrate a corporate network through a publicly accessible website, while the second demanded that the model compromise an initial network to gain unauthorized access to a separate internal system.
OpenAI, based in San Francisco, has restricted access to its Daybreak project, reflecting industry caution over releasing automated tools that can discover software vulnerabilities and exploit computer networks. However, the study found that proprietary safeguards are losing their protective lead as open-weight competitors advance.
Drastic difference in operational costs
A much larger gap emerged in the financial cost required to carry out the attacks. Generating the tokens needed to run GPT-6-Sol cost between $17 and $46 per task, whereas DeepSeek accomplished the operations for just $2 to $3.
Because open-weight models allow researchers and operators to run software locally on private hardware rather than paying for cloud programming interfaces, their low operating costs make them attractive to cybersecurity professionals testing network defenses, as well as to potential attackers.
The findings indicate that keeping top-tier models behind closed doors no longer guarantees a significant advantage in cyber warfare. In April, rival laboratory Anthropic offered limited access to its Claude Mythos Preview in an effort to buy time before similar capabilities proliferated, yet experts warned that sophisticated hacking capabilities would soon reach a wider public.
Vals noted that while cybercriminals have already started employing artificial intelligence tools, digital attackers still predominantly rely on older and less powerful systems. The firm concluded that the technical divide between open and closed systems continues to narrow, with lower costs ensuring open alternatives become ever more accessible.
