Open-Weight DeepSeek Model Nears OpenAI's Gated Cyber Model in Network Attack Tests

On Vals's CyberOffense benchmark, OpenAI's GPT-6 Sol Daybreak, which is available only to vetted partners through the company's Daybreak cyber programme, completed about 81% of attack objectives. DeepSeek V4.1 Flash 1017 completed about 73%, and Vals says open-weight models run at about a tenth of the cost per attempt.

The result goes to the heart of how the leading US labs release their most capable models. Anthropic limited the release of Claude Mythos Preview in April and uses its Project Glasswing programme to give trusted defenders early access to the full capabilities of new models, on the reasoning that gated access would buy time before advanced hacking skills spread. Vals's findings suggest the gap that strategy depends on has narrowed to a few percentage points.

"The best frontier models, locked up inside the labs and available only through trusted access programs, are now starting to be matched by open models that anybody can download," Nikil Ravi, a researcher at Vals, told Wired.

The benchmark runs each model inside the open-source Pi agent on an isolated Kali Linux machine that can reach only the model provider and the cyber range. A run ends when the agent finishes, exhausts its token budget or reaches a 48-hour limit. Vals awards credit only when a hidden marker tied to an objective appears in the recorded command output, not when a model says it has succeeded, and it counts only objectives completed within 500 million tokens so that models are compared at equal compute.

There are two ranges. Nova has 21 objectives on a large Windows network with several connected Active Directory domains, and the agent starts from a workstation exposed through a public web application. Lighthouse has 31 objectives across two separate corporate networks, where access to the first is needed to reach the second, and later stages involve phishing and abuse of trust between domains.

The two leading models split the ranges. On Nova, DeepSeek averaged 16.3 objectives over three runs against 15.7 for Daybreak. OpenAI's model built its overall lead on Lighthouse, where it completed 27 objectives to DeepSeek's 21. Other models on the leaderboard, including Kimi K3.1 Pro, Meta's Wonder and Grok 4.7, scored lower. The benchmark page was last updated on October 23.

The cost difference is large. Vals told Wired that an attack costs $17 to $46 in tokens with GPT-6 Sol, against $2 to $3 with DeepSeek; the benchmark page puts a Nova run at about $22 for Daybreak and under $3 for DeepSeek. Open models are cheaper partly because they can run on an operator's own hardware. Rayan Krishnan, Vals's chief executive, told Wired that the price gap also gives defenders a strong reason to use open models.

OpenAI did not immediately respond to Wired's request for comment.

The same DeepSeek model has already appeared in real attacks. In a report published on November 5, internet scanning firm Censys wrote that DeepSeek v4.1-flash was the main model behind an instance of Artex, an open-source AI penetration testing tool, that CrowdStrike linked to intrusions at a South Korean finance company between late September and early October.

Real-world attackers have still mostly relied on older and less capable models, both open and closed, to conserve resources, Wired noted. Defenders also have to assume that the best-funded attackers have had some access to closed models all along.

Alex Zenla, co-founder and chief technology officer of cloud security firm Edera, told Wired that she agreed open-weight models can already produce advanced attacks, especially versions whose safety training has been stripped out by modifying their weights. She said the closed labs have likely trained unreleased models that go further, and described the situation as "a cat-and-mouse game." "Unfortunately, I do think this may be the beginning of a dangerous security landscape," she said.