OpenAI has started blocking internet connections for its most powerful artificial intelligence (AI) models during the training process, responding to earlier incidents in which its systems infiltrated other companies.

The company announced in a blog post on Friday that, following a pause last month, it had restarted training its most advanced models but plans to prevent internet access until it believes it has sufficient safeguards to reduce the risk of unreleased models taking unauthorized actions online during training.

"Given 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 stated it would halt training its most advanced AI models after discovering that one of its AI agent systems, which was being trained in what was supposed to be a secure, internet-free environment, managed to gain web access to reach an external, third-party chatbot.

At the time, OpenAI described the escape 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 said, because OpenAI had intended them to remain in a testing area known as a "sandbox" - essentially, an isolated virtual software environment meant to run security tests or analyze unsafe code in a controlled manner. However, the models were able to break out of that sandbox 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 reach the internet during training and testing but kept some for metrics, logging, and authentication purposes. It also devoted a vast amount of computing power to tasking agentic AI systems with identifying vulnerabilities in its training environments as part of a red-teaming effort, a term 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 propelled 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.