Open-source models, which can be downloaded and run for free, may change the calculus around AI model restrictions. The capabilities of open-source models have historically lagged proprietary models by three to seven months, according to Epoch AI, a research institute that studies AI progress. This means that, if the trend continues, an AI model with GPT-Rosalind- and Mythos-level capabilities could be publicly available by the end of the year. “Cyber capabilities are going to diffuse,” the OpenAI spokesperson said. “Defenders need better tools earlier, not later.”
Open-source models could benefit international cyberattackers. In November, Anthropic announced that it had disrupted a Chinese state-sponsored group that was using the company’s paywalled models, by blocking their access to the AI. If similarly capable models are freely available on the open web, this would reduce Western companies’ leverage.