Open-weight artificial intelligence models have nearly caught up with the best closed systems at breaking into computer networks, according to new research from Vals, a company that measures AI capabilities. The findings suggest that restrictions the leading labs place on their most capable models may no longer preserve a meaningful advantage in cyber skills. For US companies that rely on those gated-access programs for defense, the results indicate that comparable hacking capability is now downloadable by anyone.
Vals tested how well different AI models could carry out complex, multi-step network instructions on mock corporate networks. In one scenario, models had to reach a simulated corporate network through a public-facing website. In the other, they had to compromise one company network to gain access to a second. The researchers found OpenAI's most advanced cyber model, Daybreak, which is available only to trusted partners, was the most capable, completing 81 percent of its assigned tasks on average across two tests. The Chinese open-weight model DeepSeek trailed closely, completing 73 percent of objectives on average, and the two scored equally on some tasks.
Cost emerged as a major differentiator. Vals says carrying out an attack costs $17 to $46 in tokens with GPT-6-Sol, compared with $2 to $3 with DeepSeek. Open models are generally cheaper because they can be run on a user's own hardware, and Vals CEO Rayan Krishnan said that gives defenders a strong incentive to choose open models over proprietary ones.
Researcher Nikil Ravi said the best frontier models, kept inside the labs and available only through trusted access programs, are now starting to be matched by open models anyone can download. The timing is notable because US frontier labs, particularly Anthropic, based the limited release of new models, beginning with Claude Mythos Preview in April, on the idea that gated access would only buy time before advanced hacking capabilities became broadly available. Logan Graham, who leads Anthropic's frontier red team, said in April that the company's Project Glasswing, which gives trusted partners full access to new models, aimed to put those tools in defenders' hands for a head start. He said then that the world needed to prepare for broad availability of these capabilities within 6, 12 or 24 months. Seven months later, Vals' research suggests that transition has arrived about on schedule.
Cybersecurity experts caution that the picture is mixed. Alex Zenla, cofounder and chief technology officer of cloud security firm Edera, agreed that open-weight models can produce advanced attacks and vulnerabilities today, especially modified versions that disable safety training by altering weights. Zenla cautioned that closed labs have likely trained additional, unreleased models beyond those capabilities, describing the situation as a cat-and-mouse game in which labs try to outpace open-weight models and fix vulnerabilities before those models find them. Zenla said the trend may mark the beginning of a dangerous security landscape.
In practice, real-world observation of cybercriminal and state-backed hacking in recent months has repeatedly shown attackers conserving resources and generally using older, less capable models, both proprietary and open, for most of their activity. Defenders must also assume that the best-funded and most highly motivated attackers have had at least some access to closed models all along, even with strict controls in place. Experts are still debating whether AI tools represent a new kind of threat or amplify long-standing security challenges.
More cybersecurity news from TechManNews.
