DeepSeek Finished 73% of a Simulated Corporate Breach Against Daybreak's 81%, at a Tenth of the Token Cost

Open-weight models that anyone can download are now within eight percentage points of the most capable cyber model OpenAI will not sell to the public, according to a Vals evaluation published on November 6 and reported by WIRED. The firm measured how far different models could get inside simulated corporate networks, and its numbers cut directly at the argument that gating frontier models slows the spread of offensive hacking skill.

The setup was two scenarios. In the first, a model had to work its way into a mock corporate network from a public-facing website. In the second, it had to break one company network in order to reach a second, connected one, the shape of a supply-chain intrusion. OpenAI's Daybreak, released only to vetted partners, completed 81 percent of its assigned objectives on average across the two tests. DeepSeek, whose weights can be downloaded, completed 73 percent. GLM, another open-weight family, landed close behind, and Vals notes that on some tasks the two models scored identically.

Price separated them far more than capability. Vals puts the token cost of running an attack at $17 to $46 with GPT-6-Sol, against $2 to $3 with DeepSeek, roughly a tenth, because open weights can be served on hardware the operator already owns. "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," Vals researcher Nikil Ravi told WIRED.

That comparison lands on a rationale the labs have repeated all year. Anthropic began restricting access to new models with Claude Mythos Preview in April, with its frontier red team lead Logan Graham arguing that controlled release through Project Glasswing would give defenders "6, 12, 24 months" of head start before the capability went broad. Seven months later, the gap that policy assumed has largely closed on this test.

Vals chief executive Rayan Krishnan points out that the same economics work in the other direction: a security team can probe its own estate with an open model at $2 or $3 a run instead of $17 to $46, which makes continuous automated testing affordable in a way metered API access never was.

The caveats matter. Attackers observed over the past year have mostly kept using older, cheaper models for routine work, so parity on a benchmark does not mean immediate uptake in the wild. Alex Zenla, chief technology officer at cloud security firm Edera, told WIRED that closed labs have probably trained unreleased systems beyond what Daybreak shows publicly, and that builds with safety training stripped out of the weights are already circulating. He described the dynamic as a cat-and-mouse game, and warned that this may be the start of a harder security landscape.

What the result changes is the policy argument rather than the threat level overnight. If downloadable weights reach near-parity with restricted ones within months, access control buys time at the margin and little more, and the burden shifts to whether defenders can run the same models at the same speed. OpenAI did not respond to WIRED's request for comment before publication.