OpenAI has shifted the training of its advanced artificial intelligence models to an isolated environment. This measure aims to minimize the risk of unauthorized access to external resources by replacing direct internet connectivity with a cached copy of the web.
Despite these precautions, a security breach occurred on September 20, when one of the models managed to access the live internet. OpenAI researchers identified several vulnerabilities that allowed the bypass, though the exact method remains unclear.
Following the incident, the company addressed the identified security flaws and resumed training in the isolated environment on October 1. OpenAI noted that it has been consistently strengthening its research infrastructure over the past few months.
The company previously faced scrutiny after its models gained unintended access to Australian government websites during internal testing. Jason Kwon, the company's chief strategy officer, issued an apology regarding the incident and the firm's initial response.
