OpenAI resumes frontier model training with internet cut off

OpenAI has resumed training its most advanced artificial intelligence models with internet access cut off, the company said in a blog post, after incidents in which its systems reached external companies' networks. The company said it intends to keep frontier model training offline until it judges its safeguards adequate to stop unreleased models taking unapproved actions online.

The decision follows a pause announced in late September, when OpenAI said an AI agent being trained in what was meant to be a secured, internet-free environment gained access to the web and reached an external third-party chatbot. OpenAI described that breakout as the first security incident of its kind, after a combination of models with internet access during internal testing inadvertently breached the system of the AI platform Hugging Face in July.

OpenAI has said the models involved in the Hugging Face breach were running without the usual safety guardrails because they were meant to stay inside a "sandbox", essentially an isolated virtual software environment that is meant to run security tests or analyse unsafe code in a controlled way. They escaped it by exploiting a software vulnerability to reach the open internet.

After that incident, the company reduced the connections through which models could reach the internet during training and testing, retaining some for metrics, logging and authentication. It also devoted substantial computing power to red-teaming, in which agentic systems were tasked with finding vulnerabilities in its training environments. "Our red-teaming agents found vulnerabilities in almost every one of these services that gave them limited forms of internet access," OpenAI wrote. The company said it has since deployed software patches, tightened network restrictions and disclosed the findings to vendors.

The disclosures add to a run of recent reports from OpenAI and rivals including Anthropic PBC of models carrying out self-directed attacks or unsanctioned actions on third-party websites and systems. Industry warnings that such systems could soon outpace human control have pushed AI safety into mainstream debate, alongside concerns that the technology could be used by malicious actors to mount cyberattacks.