From WIRED · Will Knight · · 1m 50s
Chinese AI models now match US labs at hacking
Open-weight AI models from China are nearly as capable as closed frontier models at breaking into networks, suggesting security restrictions may no longer slow the spread of hacking tools.
In brief
Open-weight AI models from China are nearly as capable as closed frontier models at breaking into networks, suggesting security restrictions may no longer slow the spread of hacking tools. Open-weight AI models now match closed frontier models at network hacking, undermining efforts by US labs to control access to…
Open AI models match closed labs
Vals, a company measuring AI capabilities, tested how well different models could execute complex network attacks on mock corporate networks. Open-weight models now nearly match closed frontier models.
Closed models lose their edge
Nikil Ravi, a researcher at Vals, says the best frontier models locked inside labs are now matched by open models anybody can download. This undermines efforts to limit access.
“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.”
US labs expected this timeline
Anthropic's Logan Graham told WIRED in April that the company needed to prepare for a world where hacking capabilities would be broadly available in six, twelve, or twenty-four months.
“We need to prepare now for a world where these capabilities are broadly available in 6, 12, 24 months.”
DeepSeek nearly matches Daybreak
OpenAI's Daybreak completed eighty-one percent of hacking tasks on average. The Chinese open-weight model DeepSeek completed seventy-three percent. On some tasks the two models scored equally.
Open models cost far less
Vals found it costs seventeen to forty-six dollars in tokens for GPT-six-Sol to carry out an attack compared to two to three dollars for DeepSeek. Open models run on personal hardware.
Attackers still prefer older tools
Real-world observation shows cybercriminals and state-backed hackers continue using older, less capable models to conserve resources. Best-funded attackers likely already had access to closed models.
Labs race to stay ahead
Alex Zenla, chief technology officer of Edera, says frontier labs have trained unreleased models exceeding open-weight capabilities and are racing to find vulnerabilities before open models do.
“My caveat to this is that it is likely that the closed labs have trained additional, unreleased models exceeding these capabilities.”
The main idea
Open-weight AI models now match closed frontier models at network hacking, undermining efforts by US labs to control access to advanced cyber capabilities.
