Vals finds DeepSeek completes 73% of hacking tasks, close behind OpenAI’s gated cyber model

A report on new research from Vals, a company that measures AI capabilities, indicating that open-weight models are now nearly as capable at hacking computer networks as the best closed ones. Vals tested how well different models could carry out complex, multi-step instructions on simulated corporate networks. Researcher Nikil Ravi said the leading frontier models, held back inside labs and offered only through trusted-access programmes, are starting to be matched by models anyone can download.

Two scenarios were used. One involved a mock company network reachable through a public website. The other involved two networks, where a model had to compromise the first to reach the second. OpenAI's most advanced cyber model, Daybreak, available only to trusted partners, completed 81 percent of its assigned tasks, averaged over two tests, while the Chinese open-weight model DeepSeek completed 73 percent, and the two scored equally on some tasks. Cost differed sharply: Vals says an attack by GPT-6-Sol costs 17 to 46 dollars in tokens, against 2 to 3 dollars for DeepSeek. Vals CEO Rayan Krishnan said this gives defenders a strong reason to prefer open models.

The findings bear on the strategy of US labs, notably Anthropic, which limited the release of models beginning in April with Claude Mythos Preview, on the assumption that gated access would only buy time. The article notes that real-world observation shows attackers mostly rely on older, weaker models, and that experts still disagree on whether AI changes cybersecurity fundamentally. Edera's Alex Zenla called the situation a cat-and-mouse game and said it may mark the start of a dangerous landscape. OpenAI did not immediately comment.