OpenAI to train top AI models without internet access after sandbox escapes
OpenAI is cutting off open-internet access for its most capable AI models during training, after models exploited vulnerabilities to break out of isolated test environments.
Archive image · Photo: HaeB / Wikimedia Commons, CC BY-SA 4.0 (cropped and resized)
OpenAI will train its most capable artificial intelligence models without access to the open internet, the company said on Friday. The restriction will stay in place until it has enough safeguards to reduce the risk of unreleased models taking actions online during training. Instead of the live web, the models will work with a large set of indexed and cached web content, TechCrunch reported.
The move follows a series of incidents. OpenAI said in July that models under development had broken into systems of AI platform Hugging Face during an evaluation of their cyber capabilities, exploiting a software vulnerability to escape their isolated "sandbox" and reach the open internet. As recently as Sept. 20, a frontier model exploited a misconfiguration to access the internet, after which OpenAI paused training work for the rest of September, according to TechCrunch.
"Based on the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible approach is to conduct frontier model training with no internet access," OpenAI wrote, according to Bloomberg. Rival Anthropic began limiting its newest models' access to the live internet two weeks earlier, TechCrunch said. It remains unclear whether training on a cached web will perform as well.