Safety · 2026-11-06
Open AI Models Have Almost Caught Up With Closed Ones at Breaking Into Networks
Research from Vals found that free open-weight models, including China's DeepSeek, perform complex hacking tasks almost as well as the best closed models.
What was tested
Vals, a company that measures AI capabilities, tested how well different AI models carry out complex, multi-step tasks in simulated corporate networks. These were not real systems but specially built test environments.
In one scenario, the network was reachable through a public website. In the second, there were two separate company networks, and the model had to compromise the first to gain access to the second.
Who came out on top
The most capable model was Daybreak from OpenAI, the company's most advanced cyber model, available only to trusted partners. It completed 81 percent of the tasks.
China's open-weight model DeepSeek was close behind, completing 73 percent of the tasks on average. On some tasks, the two models scored equally.
Why it matters and what it costs
According to Vals, an attack using GPT-6-Sol costs $17 to $46 in tokens, while DeepSeek costs $2 to $3. Open models are cheaper because they can run on your own hardware.
Vals CEO Rayan Krishnan believes this also gives defenders a strong reason to use open models.
What experts say
Experts see this as possibly the start of a dangerous security landscape. At the same time, real-world attackers have mostly used older, simpler models because they are cheaper. Alex Zenla of Edera noted that closed labs likely already have more powerful unreleased models.
OpenAI did not immediately respond to a request for comment.
Useful from the source
- Vals — The company that measures AI capabilities and conducted this research.
- Vals Cyber Offense benchmark — A test page showing how AI models perform cyberattack tasks in simulated networks.
Sources: 🔗 WIRED — AI