Open-weight AI models now match top closed systems in hacking prowess

New research from Vals indicates that publicly accessible, open-weight AI models are achieving hacking capabilities comparable to the most restricted, closed frontier models, signaling a major shift in cybersecurity. Nikil Ravi, a researcher at Vals, noted that open models are now matching the performance of proprietary models locked within labs. In testing across two simulated corporate network scenarios, Vals found OpenAI's proprietary Daybreak model performed best, completing 81% of tasks on average. However, the Chinese open-weight model DeepSeek was close behind, achieving 73% completion. Furthermore, Vals highlighted a cost advantage, showing DeepSeek costs only $2 to $3 in tokens compared to $17 to $46 for GPT-6-Sol. This development challenges the strategy of US frontier labs, like Anthropic, which previously argued that limited access to advanced models would buy time. Logan Graham, Anthropic's frontier red team lead, had previously stated in April that preparation for broad availability within 6 to 24 months was necessary. Alex Zenla, cofounder and CTO of Edera, agreed that open models pose a threat, but cautioned that closed labs may possess even more advanced, unreleased models, framing the situation as a cat-and-mouse game.