The ChatGPT maker said in a blog post it had been testing the capabilities of some of its most advanced models – including those not yet released to the public – in a controlled environment.
But they managed to escape containment, reach the internet, and break into a firm called Hugging Face to satisfy its testing objectives.
Hugging Face is a platform used to host open-source large language models and datasets.
Open-source models are ones which make their source code and training data available to the public for testing and study. Popular mainstream models like ChatGPT and Google’s Gemini are not open-source.
In its own blog post last week, Hugging Face said it was investigating a hack that “was different from anything we had handled before” and had been “driven, end to end, by an autonomous AI agent system”.
OpenAI took responsibility in a post on Tuesday, saying its models had carried out “an unprecedented cyber incident, involving state-of-the-art cyber capabilities”.
It was driven by a combination of models, including a next-generation version of GPT called 5.6 Sol and an “even more capable pre-released model”.
They became “hyperfocused” on achieving their testing objectives and identified the Hugging Face platform as a possible source of solutions.
OpenAI, led by Sam Altman and valued at more than $850bn, said it would reinforce its safeguards. But it also stressed the models had escaped from what it described as a “highly isolated environment”.
OpenAI and Hugging Face are conducting a joint investigation into how it happened, the blog post added.
The incident will intensify concerns over the power of the world’s most powerful AI models and the associated risks.
Hugging Face CEO said it showed the issue of AI safety “won’t be solved by any single company working in secret”, reiterating the need for collaboration between leading developers.
