Open-weight artificial intelligence models have reached a level of capability in hacking computer networks that nearly matches the industry's most advanced closed models, according to research from Vals.
The company tested various models by tasking them with complex, multi-step instructions on mock corporate networks. The study found that OpenAI’s Daybreak, a closed model available only to trusted partners, completed 81 percent of tasks. The Chinese open-weight model DeepSeek completed 73 percent of objectives on average, with researchers noting the two models scored equally on certain tasks.
“The best frontier models, locked up inside the labs and available only through trusted access programs, are now starting to be matched by open models that anybody can download,” Nikil Ravi, a researcher at Vals, said.
Beyond performance, the research highlighted significant cost differences between the two types of models. Vals reported that an attack using GPT-6-Sol cost between $17 and $46 in tokens, while the same task using DeepSeek cost between $2 and $3. Rayan Krishnan, CEO of Vals, stated that the lower cost of open models, which can be run on private hardware, provides an incentive for defenders to use them over proprietary options.
The findings suggest that efforts by frontier labs to limit access to new models may be ineffective at slowing the proliferation of advanced hacking capabilities. This development follows previous warnings from industry leaders regarding the speed of AI adoption. Anthropic frontier red team lead Logan Graham previously told WIRED that the company's Project Glasswing aimed to provide trusted access to new tools to give defenders a head start.
While AI is increasingly used to automate core hacking tasks and find vulnerabilities in code, experts continue to debate whether the technology fundamentally alters the foundation of cybersecurity or simply amplifies existing challenges. Recent observations of cybercriminal and state-backed activity show that many attackers still prefer to conserve resources by using older, less capable models for most operations.
