DeepSeek matches OpenAI's locked cyber model at a tenth of the cost: Reveneau

Why it mattersThe cost gap between open and closed AI models for offensive security is now ten to one. Defenders can run the same tool.

The strongest AI tool for hacking a network used to be locked inside OpenAI. Now a free Chinese model scored level with it. Wired reports the test. Vals.ai, a company that measures AI models, ran top AI models through multi-step network hacking tasks. The headline finding is close scores for DeepSeek V4.1 Flash against OpenAI's restricted GPT-6 Sol Daybreak.

Vals.ai publishes the test as its CyberOffense leaderboard. The benchmark uses two simulated corporate networks. Nova is a single Windows network with linked Active Directory domains. Lighthouse is harder. It has two separate networks. The model has to breach the first network, take credentials, and use them to breach the second.

The numbers Vals.ai publishes

On Nova, DeepSeek V4.1 Flash scored 16.3 of 21 objectives. GPT-6 Sol Daybreak scored 15.7 of 21. On Lighthouse, DeepSeek scored 21 of 31 and Daybreak scored 27 of 31. Z.ai's GLM 5.3 Flash scored 5 of 21 and 5 of 31 on the same tests.

Vals.ai also publishes the token cost per run. DeepSeek V4.1 Flash ran the first scenario for under $3. GPT-6 Sol Daybreak ran it for about $22. Wired quotes a wider range of $17 to $46 for GPT-6 Sol across the full set. Open models are cheaper because a user can run them on their own hardware, Vals.ai notes.

Who said what

Nikil Ravi, a Vals researcher, spoke to Wired. He said the best frontier models are now being matched by open models anybody can download. Rayan Krishnan, Vals.ai's CEO, told Wired defenders now have a reason to use open models instead of proprietary ones.

Anthropic set this up earlier in the year. Logan Graham is its frontier red team lead. He told Wired the trusted-access program for Claude Mythos Preview was meant to give defenders a head start. Graham said the window was 6 to 24 months. Vals's numbers suggest the window closed faster than that.

Alex Zenla, chief technology officer at Edera, told Wired he agrees open-weight models can produce advanced attacks today. He added a caveat. Closed labs probably have stronger unreleased models. The labs stay ahead by training new ones faster than open ones catch up. OpenAI did not respond to Wired's request for comment.

What a software team with real attackers in the threat model takes from this

Two things change on paper, and one in practice. First, the tool asymmetry is smaller than it was a year ago. The best offensive AI tool is still restricted. The next tier down is a download. Second, cost is no longer a gate on volume attacks. A Daybreak-level run on 1,000 targets used to cost $17,000 to $46,000 in tokens. The same run on DeepSeek costs $2,000 to $3,000. It scales on cheaper hardware.

The third thing is what defenders do with the same model. Krishnan's point is a serious one. A blue team can run DeepSeek V4.1 Flash against its own networks. It can use Vals-style multi-step scenarios. It gets a close picture of what a well-funded attacker can now do. The price fits in a quarterly budget. That is the move to make this quarter. The attackers already have the model.

Source

Free Chinese AI Models Are Getting Scary Good at Hacking, Wired, by Will Knight and Lily Hay Newman
CyberOffense Leaderboard and Methodology, Vals.ai

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.