When was OpenAI introduced? the first models with an open scale in years, not only technology companies drew attention. The release also excited U.S. military and defense contractors, who saw an opportunity to exploit them for highly secure operations.
Preliminary results show that OpenAI tools are lagging behind the competition in terms of desired capabilities, some military suppliers tell WIRED. But they are still cheerful that models from a key industry leader are finally an option for them.
Lilt, an artificial intelligence translation company, contracts with the US military to analyze foreign intelligence data. Because the company’s software handles sensitive information, it must be installed on government servers and run without an Internet connection, a practice known as air-gapping. Lilt has previously developed its own AI models or used open source options such as Meta’s Llama and Google’s Gemma. However, OpenAI tools were out of the question because they were closed source and could only be accessed online.
The creator of ChatGPT’s fresh open source models, gpt-oss-120b and gpt-oss-20b, have changed that. Both can run locally, which means users have the freedom to install them on their devices without having to connect to the cloud. And with access to model weights – key parameters that determine their response to various prompts – users can tailor them to specific goals.
OpenAI’s return to the open source market could ultimately augment competition and lead to better-performing systems for the military, healthcare companies and others working with sensitive data. In recent McKinsey study of approximately 700 business leaders, more than 50 percent said their organizations exploit open-source AI technologies. Models have different strengths depending on how they are trained, and organizations often exploit several together, including open weight models, to ensure reliability in a variety of situations.
Doug Matty, director of digital and artificial intelligence at the so-called War Department, which the Trump administration uses for the Department of Defense, tells WIRED that the Pentagon plans to integrate generative artificial intelligence into battlefield systems and back-office functions such as auditing. Some of these applications will require non-cloud models, he says. “Our capabilities need to be flexible and adaptable,” says Matty.
OpenAI did not respond to requests for comment on how its open-source models might be used by the defense industry. Last year, the company lifted a broad ban on the exploit of its technology for military and war applications, prompting criticism from activists concerned about the harms caused by artificial intelligence.
In the case of OpenAI, offering a free and open model can have several benefits. Ease of access could cultivate a larger community of experts in their technologies. And because users don’t have to register as formal customers, they can be able to operate in secret, which could protect OpenAI from criticism from potentially controversial customers such as the military.