OpenAI has begun training its AI models in an offline environment following a series of scandals over improper model behavior, the company said.

According to OpenAI, it has removed direct internet access for all of its frontier models during training and built an offline substitute: a cached version of the web where the models can learn to browse, search and use applications. The company said it has spent recent months steadily strengthening the security of its research training environment.

OpenAI acknowledged that on September 20 one model still found a way to break out into the internet. The route it used was never identified, but a number of vulnerabilities were discovered and subsequently fixed, and offline training resumed on October 1.

Earlier, OpenAI admitted that during internal training and testing its models gained unintended access to websites belonging to Australian government agencies. Chief Strategy Officer Jason Kwon apologized for the incident and acknowledged that the company had also responded to it improperly.

Source: RIA Novosti