Free AI Models Now Rival Locked-Up Ones at Hacking Tests

Free AI models from China are now close to the best locked-up ones at breaking into computer networks. That is the finding of a new test by Vals, a company that measures what AI can do. Wired reported the results. The gap between models only trusted partners can use and models anyone can download is shrinking.

What the test found

Vals set up mock company networks and asked AI models to complete multi-step hacking tasks. One scenario had a network reachable through a public website. The other had two company networks, where a model had to break into the first to reach the second.

• OpenAI's Daybreak, open only to trusted partners, completed 81 percent of tasks across two tests.
• DeepSeek, a Chinese open-weight model, completed 73 percent on average.
• On some tasks, the two scored equally.
• Wired's summary says another Chinese model, GLM, also landed not far behind.

Open-weight means the model's inner settings are public, so anyone can download it and run it on their own computers.

Why cost matters

Vals says an attack cost $17 to $46 in tokens with GPT-6-Sol, an OpenAI model. With DeepSeek it cost $2 to $3. Tokens are the small chunks of text AI companies bill for. Open models tend to be cheaper because they can run on your own hardware. Vals CEO Rayan Krishnan says that gives people playing defense a strong reason to use open models too.

Why labs are watching

US labs, especially Anthropic, limited the release of new models starting with Claude Mythos Preview in April. The idea was that restricted access would only buy time. Anthropic's Logan Graham said in April the world should prepare for these abilities to be widely available in 6, 12 or 24 months. Seven months later, Wired says, that shift seems to have arrived on schedule.

What is still unclear

The picture is not all alarming. Wired reports that real attackers have mostly stuck with older, less capable models, both open and proprietary, to save resources. Experts also still debate whether AI creates a new kind of threat or simply speeds up old problems.

Alex Zenla of the security firm Edera calls it a cat-and-mouse game. She says closed labs have likely trained unreleased models that are even stronger. She also warns this may be the start of a dangerous security landscape. OpenAI did not immediately respond to a request for comment.

What to do

Cheaper hacking tools mean more people can try attacks. The basics matter more, for a business and for a family.

• Install software updates soon after they appear.
• Use strong, unique passwords.
• Turn on two-step login wherever it is offered.