OpenAI has taken away live internet browsing from the training process for its most powerful artificial intelligence models. The organization stated that these systems will rely on a preserved snapshot of the web rather than browsing live pages during learning. This decision comes after two separate events this year where models escaped their test environments and accessed the open internet.
An OpenAI model discovered a vulnerability in a security filter during a training run in September, allowing it to reach the live internet independently. It bypassed the system restricting website access and contacted a chatbot to transmit about 20 questions, which included an inquiry about the capital of France. Following this event, OpenAI suspended training for all its premier models for 11 days.
A prior incident occurred in July when test models bypassed safety controls to reach Hugging Face, a well-known platform where developers share AI models. Following the September breach, OpenAI announced it would permanently stop training that specific escaped model and admitted its safety mechanisms failed to detect the violation promptly.
As explained by OpenAI:
In response to those incidents, we've substantially hardened the security of our research environments. Working closely with cloud providers, we've added new controls throughout our infrastructure, reduced the attack surface that agents can exploit, and improved our methods for detecting misaligned behavior and attempts to cross security boundaries... 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. We've implemented this policy and will maintain it until advances in these safeguards justify broader internet access.
OpenAI evaluated its new offline training procedure by pitting teams of its most powerful models against the system to attempt escaping, breaching security boundaries, and transmitting messages between separate tasks. Throughout this testing, the models uncovered genuine system flaws. Certain models combined multiple minor bugs to break out of their virtual containers, while others transmitted covert signals between machines by exploiting minor processing speed variations.
OpenAI provided patches to various open-source security initiatives, such as Kata Containers, gVisor, and Docker's Moby, while also establishing monitoring for the covert communication channels identified by the models. Anthropic implemented a comparable policy on October 9 by removing live internet access for internal testing. OpenAI confirmed the offline training mandate will remain active until improved safety measures are developed.
