Open-weight DeepSeek nears Daybreak on hacking tasks, Vals finds

The Vals experiment lands as frontier labs, particularly Anthropic, have argued that gated access to advanced models can buy time before hacking capabilities become widely available. Anthropic's Logan Graham told WIRED in April that the company's Project Glasswing aimed to get tools to defenders for a head start, and that the world needed to prepare for broad availability in 6, 12, or 24 months. Seven months later, Vals's findings suggest that transition has arrived on schedule, with open-weight models like DeepSeek matching restricted ones on many simulated tasks.

Yet real-world observation of cybercriminal and state-backed activity has repeatedly shown attackers conserving resources and making do with older, less capable models for most of their activity. Even with strict controls, defenders must assume the best-funded attackers have already had some access to closed models. Alex Zenla of Edera cautions that closed labs likely have additional unreleased models exceeding these capabilities, framing the situation as a cat-and-mouse game in which labs race to secure vulnerabilities before open-weight models find them.