OpenAI disables internet access during model training

OpenAI has announced that it will deny internet access during the training of its frontier models. This decision underscores the inadequacy of the security measures implemented thus far. The company stated in a blog post that, based on the current state of alignment research and monitoring systems, it believes that conducting training without internet access is the responsible approach.

This policy comes as a response to a report from safety evaluation partners METR and Redwood Research, which found OpenAI's security and monitoring measures over the summer to be inadequate. In light of this, the company has claimed to have substantially hardened the security of its research environments and implemented new controls over its infrastructure.

Why it matters

OpenAI's decision to restrict internet access during model training reflects growing concerns about security in AI development. This move not only aims to protect the integrity of the models but also highlights the need for improved security measures within the AI industry.

What we know

OpenAI has stated that it will maintain this restricted access policy until advances in security safeguards justify broader internet access. This indicates the company's commitment to prioritizing safety in its model development and training processes.

What remains unclear

It has not yet been confirmed when these advancements in security measures that would allow for broader internet access during training are expected to be achieved. It also remains to be seen how this policy will affect the development of future AI applications by OpenAI.