Open-weight artificial intelligence (AI) models are now nearing the capabilities of some of the most sophisticated closed systems in terms of hacking into computer networks, according to recent research that marks a significant shift in the cybersecurity landscape. The study, conducted by Vals—a firm specializing in evaluating AI capabilities—focuses on how well various AI models can execute complex, multi-step instructions across simulated corporate networks.
The research sheds light on a concerning trend in cybersecurity: open models, which are freely available for download, are approaching the performance levels of elite models that are restricted and only accessible through vetted programs. Nikil Ravi, a researcher at Vals, indicates that the best frontier models, long confined within laboratory walls, are now being closely matched by models that can be utilized by anyone with access to the internet. This realization calls into question the effectiveness of efforts undertaken by leading laboratories to restrict access to their advanced models in an attempt to curb the spread of such hacking capabilities.
The findings highlight a pivotal moment for organizations that rely on cyber defenses, especially in light of recent strategies adopted by U.S. research labs—most notably Anthropic. In April, Anthropic initiated limited release of new models under the premise that restricting access would provide a buffer against the widespread availability of advanced hacking techniques. According to Logan Graham, Anthropic’s red team lead, the company’s Project Glasswing aims to equip defenders with the necessary tools to preemptively counteract emerging threats.
As the research findings unfold, the anticipated transition to a new era of accessible cyber capabilities appears to be well underway. Vals conducted experiments using AI in two different simulated scenarios, assigning them particular tasks within a mock corporate environment. The first scenario evaluated a model’s ability to breach a network accessible via a public-facing website, while the second required the model to compromise one network to gain access to another.
Remarkably, Vals discovered that OpenAI’s advanced cyber model, Daybreak, which is available exclusively to trusted partners, achieved an 81% success rate in completing designated tasks across two evaluations. Close behind was the open-weight Chinese model DeepSeek, which accomplished approximately 73% of the tasks. The researchers noted that, in some scenarios, both models performed similarly, suggesting a narrowing gap between open and closed model capacities.
The economic implications of using open-weight models were equally striking. Vals indicated that the cost per attack for GPT-6-Sol, a proprietary model, ranged from $17 to $46, while DeepSeek’s expenses were substantially lower, between $2 and $3. This affordability is largely due to the ability to operate open models on personal hardware, making them a tempting option for many organizations, including those focused on defensive strategies.
As this technology continues to evolve, business owners must remain vigilant about the potential threat posed by accessible AI-driven cyber tools. With sophisticated tactics now seemingly within reach of a wider array of individuals, the spectrum of adversarial methods could include initial access techniques, persistence strategies, and privilege escalation actions, as classified under the MITRE ATT&CK framework. The implications for cybersecurity defenses are significant, as organizations must reassess their strategies in light of these emerging capabilities and the inevitable proliferation of advanced hacking techniques.