Free Chinese AI Models Nearly Catch OpenAI’s Best Model at Hacking

A Vals study found that the open Chinese model DeepSeek completed 73% of tasks in simulated corporate network hacking scenarios, compared with 81% for OpenAI’s closed Daybreak. DeepSeek was also significantly cheaper to use.

Free Chinese AI models nearly catch OpenAI’s best model at hacking
AI-generated illustration

What the comparison found

The open Chinese model DeepSeek nearly matched OpenAI’s closed system in a test of multi-step cybersecurity tasks. The study was conducted by Vals, a company that evaluates AI capabilities. The test used two simulated corporate scenarios, not real networks.

In the first scenario, the model had to work with a hypothetical network accessible through a public website. In the second, it had to gain access to one simulated network and then move to another. This format tests more than an individual command: it assesses whether AI can solve a connected task step by step.

OpenAI’s Daybreak had the best result, completing an average of 81% of the tasks. It is available only to trusted partners. DeepSeek completed 73% of the objectives. Vals researchers note that the models tied on some individual tasks.

The open model costs far less

The gap was not just in performance. Vals estimates that carrying out an attack with GPT-6-Sol costs $17–46 in tokens, compared with $2–3 for DeepSeek. Open models can run on your own hardware, so using them does not depend on access to a closed platform or its pricing.

That shifts the balance of capabilities in cybersecurity. Until recently, the most capable models were available mainly through programs with limited access: labs sought to limit the spread of dangerous skills and give defenders time to prepare. Now, comparable capabilities are gradually emerging in systems that can be downloaded and run independently.

That does not mean that any given attacker is already using these models for sophisticated attacks. Observations of real-world cybercriminal and state-backed activity show that attackers more often conserve resources and use older or less capable systems. But lower costs and the availability of open weights broaden the pool of potential users.

Cloud security expert Alex Zenla believes open models can already develop sophisticated attacks and find vulnerabilities. At the same time, closed labs have likely trained even more powerful systems that they have not yet released. As a result, the race between open and closed models continues, and the line between tools for attack and defense is becoming less distinct.

Learn more

Original publications for a closer look.

Open Chinese AI Models Are Getting Scary-Good at Hackingwww.wired.com
← All articlesMain website ↗

Have a process to automate?

We turn ideas into working products, from the first workflow to launch.

Discuss a project