Vals: open-weight DeepSeek nears OpenAI's restricted cyber model in simulated network attacks

Vals AI let AI agents attack two simulated company networks on their own, and the best result came from OpenAI's GPT-6 Sol, which is available only through OpenAI's restricted Daybreak cyber program, according to Vals' new cyber range benchmark. The open-weight model DeepSeek V4.1 Flash 1017, which anyone can download, came close behind.

On the harder range, called Lighthouse, GPT-6 Sol Daybreak broke into the first company through a public web application, took administrator control, stole passwords and used them to reach a second company, completing 27 of 31 objectives in under three hours. DeepSeek scored level with Daybreak on the first network for under $3 per run, against about $22 for Daybreak.

The two models used compute very differently. GPT-6 Sol reached 25 of 31 objectives within 100 million tokens, while DeepSeek completed nothing in its first 25 million tokens and made most of its progress after 100 million, spending two to four times as many tokens. Grok 4.7, Kimi K3 and GLM 5.3 Flash stalled well before their budgets ran out, and no model completed the final objectives on the two-network range.

"Once a model's weights are public, no provider can monitor or stop how it is used," Vals wrote. It said the same models can help defenders find weaknesses first, and that it expects attacks of this kind on real enterprise networks to become more common in the coming months.