OpenAI Trains Its Models on an Offline Internet

OpenAI restricts web access for its developing models and shifts training, like most evaluations, to an offline internet. The company highlights major risks associated with the web and acknowledges the ineffectiveness of previous isolation methods. Its "offline world" relies on indexed and cached pages to limit any impact on real systems.

OpenAI Tightens Isolation and Acknowledges the Limits of Previous Safeguards

OpenAI states that its previous approaches to prevent undesirable behaviors during the development phase are ineffective. Consequently, the company is reinforcing the isolation of its models during training and evaluations by excluding online Internet from the authorized scope for most tests. It argues that web access exposes significant risks while recognizing that such access can help models acquire useful skills. OpenAI also emphasizes the risk that a high-performing, misaligned agent could influence systems beyond its control. These measures are part of a broader isolation framework announced by the company.

A Cached and Indexed Web as Training Ground

OpenAI has developed an offline version of the Internet dedicated to research and testing. This offline environment gathers indexed web pages as well as cached copies. Models are trained to explore, extract information to formulate responses, and manipulate applications, all while remaining confined to this framework. This setup is presented as a way to avoid real-world consequences. Developing models thus have very limited access to the public web and primarily rely on this closed infrastructure.

Decisions Made in the Context of Incidents and Unpredictability

These decisions follow security issues that occurred with developing AI models, notably the hacking of Hugging Face by a group of AI agents. The development period is considered a time when an AI can exhibit unexpected behaviors. It is in this context that OpenAI is reinforcing the isolation of its models.