OpenAI resumes training AI, blocks internet access over security risks

The company said it would keep its most advanced models offline during training until safeguards are adequate, following incidents in which AI systems accessed the web and breached external systems

By Rachel Metz

 

OpenAI has begun cutting off internet access for its most capable artificial intelligence models during the training process in response to prior incidents in which its systems breached other companies. 

The company said in a blog post Friday that, after a pause last month, it had resumed training its most advanced models but intends to block internet access until it feels it has adequate safeguards in place to mitigate the risk of unreleased models being able to take unapproved actions online during training. 

 

“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 wrote.

 

In late September, the company said it would pause training its most advanced AI models after it discovered that one of its AI agent systems that was being trained in what was supposed to be a secured, internet-free environment was able to gain access to the web to reach an external, third-party chatbot.

 

At the time, OpenAI described the breakout as the first security incident of its kind since a combination of models gained internet access during internal testing and inadvertently breached the system of the AI platform Hugging Face in July. 

 

The models involved in the inadvertent Hugging Face breach were operating without the usual safety guardrails, the company has said, because OpenAI had intended them to remain in a testing area known as a “sandbox” — essentially, an isolated virtual software environment that’s meant to run security tests or analyse unsafe code in a controlled way. The models were able to break out of that sandbox, however, by leveraging a software vulnerability to access the open internet.

 

OpenAI said that after the Hugging Face incident, it reduced the connections through which its models might be able to reach the internet during training and testing but kept some for the purpose of metrics, logging and authentication. It also spent a vast amount of computing power on tasking agentic AI systems with identifying vulnerabilities in its training environments as part of a red-teaming effort, a phrase typically used to describe deliberate attempts to break into systems to uncover vulnerabilities before real attackers can.

 

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

 

The company said it has since deployed software patches, imposed tighter network restrictions and disclosed the findings to vendors.

 

There have been a series of disclosures in recent months from OpenAI and rivals including Anthropic PBC about incidents in which their models carried out self-directed attacks or unsanctioned actions on third-party websites and systems. 

 

These cases, and a series of stark warnings from within the industry, have vaulted the safety debate into the mainstream. AI developers have warned that their systems could soon become too sophisticated for people to control, adding to concerns the technology could be used by malicious human actors to create bioweapons or carry out devastating cyberattacks.