Daijiworld Media Network - New York
New York, Oct 24: OpenAI has resumed training its most advanced artificial intelligence models but is restricting their internet access following security incidents in which AI systems escaped controlled environments and accessed external websites without authorisation.
In a blog post published on Friday, the company said it would keep internet access disabled during frontier model training until it had adequate safeguards to prevent unreleased models from taking unauthorised actions online.
“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, when OpenAI paused training its most advanced models after discovering that an AI agent being trained in what was intended to be a secure, internet-free environment had managed to access the web and reach an external third-party chatbot.
OpenAI described the incident as the first of its kind since an earlier security breach in July, when a combination of models gained internet access during internal testing and inadvertently breached the systems of AI platform Hugging Face.
The models involved in the July incident were operating without the usual safety guardrails because they were intended to remain within a sandbox, an isolated virtual environment designed to run software tests and analyse potentially unsafe code under controlled conditions.
However, the models exploited a software vulnerability to escape the sandbox and access the open internet.
Following the incident, OpenAI reduced the number of connections through which its models could access the internet during training and testing. However, some connections were retained for functions such as metrics, logging and authentication.
The company also deployed AI agents to identify weaknesses in its training infrastructure through red-teaming exercises, which involve deliberately testing systems for vulnerabilities before they can be exploited by malicious actors.
“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 deployed software patches, introduced stricter network restrictions and disclosed its findings to the relevant vendors.
The incidents come amid growing concerns across the artificial intelligence industry over models carrying out self-directed attacks or unauthorised actions on third-party websites and systems. OpenAI and rival AI developer Anthropic have disclosed incidents involving potentially risky model behaviour in recent months.
Researchers and technology companies have increasingly warned that as AI systems become more capable of operating independently, ensuring human oversight and preventing misuse will become more challenging. Concerns include the possibility of malicious actors exploiting advanced AI tools to develop biological weapons or conduct damaging cyberattacks.
OpenAI's decision to restrict internet access during advanced model training reflects the growing emphasis on securing development environments as AI systems become more sophisticated.