Open-weight artificial intelligence models are now approaching the cybersecurity capabilities of the best closed models, according to research from Vals. The findings suggest that restricted access may no longer provide a decisive advantage in advanced cyber skills, while the lower cost of open models could make them attractive to both attackers and defenders.
Open models close the cybersecurity gap
Vals, a company that measures AI capabilities, tested how well different AI models could execute complex, multi-step network instructions on simulated corporate networks.
“The best frontier models, locked away in labs and available only through trusted access programs, are now beginning to rival open models that anyone can download,” says Vals researcher Nikil Ravi.
The results indicate that closed models no longer hold a decisive lead in cyber skills. That could undermine recent efforts by Frontier Labs to restrict access to advanced models in an attempt to slow the spread of their capabilities.
The findings are also significant because Frontier Laboratories in the US, particularly Anthropic, had assumed that limited releases of new models would provide a temporary advantage. That approach began with the Claude Mythos Preview in April, based on the idea that controlled access would buy time until advanced hacking capabilities became widely available.
Anthropic expected a limited window
Logan Graham, Anthropic’s Frontier Red team leader, told WIRED in April that the company’s Project Glasswing, which provides trusted access to the full capabilities of new models, was intended to give defenders an early advantage.
“We need to prepare now for a world where these capabilities will be widely available in six, 12, 24 months. A lot of things are going to change when it comes to security,” he told WIRED at the time.
Seven months later, that transition appears to have arrived on schedule.
How Vals tested the models
Vals evaluated the AI models in two simulated scenarios, assigning them a range of specific tasks. The first scenario involved a simulated corporate network accessible through a public website. The second included two separate corporate networks, requiring the model to compromise the first network before accessing the second.
OpenAI’s cutting-edge cyber model, Daybreak, which is available only to trusted partners, performed best. It completed 81% of the assigned tasks across an average of two tests.
DeepSeek, an open model from China, followed with an average success rate of 73%. Vals researchers noted that the two models achieved comparable scores on some tasks.
Open models cost less to deploy
The study also highlighted the cost advantage of using an open-weight model for this type of work. Vals estimated that GPT-6-Sol tokens cost between $17 and $46 to carry out an attack, compared with $2 to $3 for DeepSeek.
Open models are generally less expensive because they can run on an organization’s own hardware. Vals CEO Rayan Krishnan said that cost difference provides a strong incentive for defensive teams to use an open model instead of a proprietary system.
Source: www.wired.com