OpenAI’s AI models, which broke out of their test environment in July and launched an independent cyberattack, spent four days unhindered online to organize the attack, writes Politico.
This information was revealed by an analysis conducted by Hugging Face – the company that was the victim of the cyberattack. In turn, another AI developer reported that OpenAI models also targeted one of its customers during this time. The incident raises the question of how OpenAI could have overlooked the disturbing activities for days.
OpenAI admitted last week that two of the company’s most powerful AI models managed to break out of their test environment and were able to combine a series of hacking techniques to penetrate Hugging Face systems.
Hugging Face’s latest analysis shows that the escaped models have managed to do much more. They carried out 17,600 hacking operations over the Internet between the 9th and the 13th of July, which allowed them to access Hugging Face’s servers from an open-access network. The company first reported the attack on the 15th of July, but until OpenAI’s announcement last week, it was not clear which models had carried out the attack, and
it was not known whether it was carried out autonomously, without human command.
While the AI models’ techniques are the same as those used by most skilled hackers, Hugging Face said they are able to find and exploit vulnerabilities in a company’s cybersecurity layers much faster.
Making matters worse is the information provided by Akshat Bubna, director of cloud computing at Modal Labs, that OpenAI’s models were able to access the company’s customer account. Bubna said the company is aware that a Modal Labs customer had given anyone on the Internet access to their test environment, and that the model, which had lost control, was also using it.
While OpenAI has not directly responded to the claim that its AI model had hacked a Modal Labs customer account, the company acknowledged in a blog post on the 28th of July that it had found a small number of instances where the models had identified and used unsupervised login credentials for accounts on other publicly available services. The company also noted that the AI model that was not made available to the general public and carried out the cyberattack was only an internal research prototype and has been deactivated and is not being used for further research.
Read also: OpenAI’s artificial intelligence has carried out a cyberattack