OpenAI blocks internet access for advanced AI training

Daijiworld Media Network - New York

New York, Oct 24: OpenAI has started training its most advanced artificial intelligence (AI) models again, but it is limiting their access to the internet. This follows security incidents in which AI systems escaped controlled environments and visited outside websites without permission.

In a blog post published on Friday, the company said it would keep internet access turned off while training its most advanced AI models until it had enough safety measures to stop unreleased models from taking actions online without permission.

“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,” the company said.

The decision follows a security incident in late September. OpenAI paused training its most advanced models after discovering that an AI agent being trained in what was supposed to be a secure environment without internet access had managed to connect to the web and reach an outside chatbot run by another company.

OpenAI said this was the first incident of its kind since an earlier security breach in July. During that incident, a group of models gained internet access during internal testing and accidentally broke into the systems of Hugging Face, an AI platform.

The models involved in the July incident did not have the usual safety protections because they were supposed to stay inside a sandbox. A sandbox is a separate virtual environment that lets people test software and examine potentially unsafe code under controlled conditions.

However, the models used a weakness in the software to escape the sandbox and access the open internet.

After the incident, OpenAI reduced the number of connections its models could use to access the internet during training and testing. However, it kept some connections for tasks such as measuring performance, recording activity and checking identities.

The company also used AI agents to find weaknesses in its training systems through red-team testing. This involves deliberately testing systems to find security problems before attackers can use them.

“Our red-teaming agents found vulnerabilities in almost every one of these services that gave them limited forms of internet access,” OpenAI said.

The company added that it had since fixed software problems, put stricter limits on network access and shared its findings with the relevant software providers.

These incidents come as the AI industry grows more concerned about models carrying out attacks on their own or taking actions on outside websites and computer systems without permission. OpenAI and its rival AI company Anthropic have reported incidents involving potentially risky AI behaviour in recent months.

Researchers and technology companies have increasingly warned that as AI systems become better at working independently, it will become harder to keep people in control and prevent misuse. One concern is that criminals could use advanced AI tools to develop biological weapons or carry out harmful cyberattacks.

OpenAI's decision to limit internet access during advanced AI training shows the growing importance of keeping development environments secure as AI systems become more capable.