DeepSeek closes the gap with OpenAI's locked-down hacking model

AIVIO News reported earlier that OpenAI had launched GPT-6 Cyber, a model built to find and fix security flaws. New testing published by the AI evaluation firm Vals suggests the distance between OpenAI’s strongest cyber-offense tools and freely downloadable rivals is shrinking faster than expected.

Vals built two mock corporate network scenarios and assigned AI models chained, multi-step intrusion tasks of the kind a hacker uses to break into a system. One scenario involved a single network reachable through a public website; the other required compromising one company’s network to reach a second, connected one.

OpenAI’s most advanced cyber model, called Daybreak and available only to trusted partners, completed 81 percent of its assigned tasks on average across two test runs, Vals found. DeepSeek, an open-weight model built in China that anyone can download, completed 73 percent of the same tasks, and Vals researchers said the two models scored equally on some individual tasks.

Vals also compared cost. Running an attack with OpenAI’s GPT-6-Sol model cost $17 to $46 in computing tokens, the firm says, against $2 to $3 for DeepSeek. Vals chief executive Rayan Krishnan said that price gap gives network defenders, not just attackers, a reason to rely on open models, since they can be run on a company’s own hardware rather than paid for per use.

Nikil Ravi, a researcher at Vals, said frontier models once available only through trusted-access programs are now being matched by systems anyone can download. That finding touches on a bet laid out by Anthropic, whose frontier red-team lead, Logan Graham, told Wired in April that restricting early access to its Claude Mythos Preview model was meant to buy defenders time, on a horizon he put at months rather than years, before similar capabilities spread widely.

Researchers tracking real-world hacking activity note that criminal and state-backed attackers have mostly stuck with older, cheaper models so far, open or proprietary, rather than rushing to adopt the newest tools. Defenders are still advised to assume well-funded attackers already had some access to closed models regardless of official gating.

What this means

The Vals results undercut part of the logic behind limiting access to the most capable cyber models: if a downloadable system like DeepSeek can approach the performance of a gated one like Daybreak, restricting distribution buys less time than labs such as OpenAI and Anthropic had hoped. The cost difference compounds that shift, since a few dollars per attack run lowers the bar for both attackers scaling up operations and defenders screening their own networks.

What the reports do not establish is whether this narrowing holds beyond the two scenarios Vals built, or against the most sophisticated state-backed operations. Alex Zenla, chief technology officer at the cloud security firm Edera, described the pattern as a cat-and-mouse game, saying closed labs likely hold unreleased models more capable than anything tested, built to stay ahead of what open systems can already do.

Zenla said the convergence may mark the start of a harder security problem for defenders, even as experts continue to debate whether AI has changed the basic shape of cybersecurity or simply sped up challenges that already existed.