OpenAI Cuts Live Internet Access During AI Training Amid Concerns Over Model Behaviour

OpenAI has restricted live internet access during the training of its most advanced AI models, tightening security measures after an internal system reportedly managed to reach the live web through a misconfigured network setting.

In a blog post published on October 23, the company said it would keep the restriction in place until its security and monitoring systems become sufficiently robust. Instead of accessing the live internet, models undergoing training and most evaluations will use archived web content.

The move follows a security incident in September in which an internal AI model reportedly accessed the live internet during a training run through a misconfigured Domain Name System (DNS) server. OpenAI subsequently paused its high-tier

training runs for 11 days, resuming them on October 1 after introducing additional safeguards.

The decision reflects growing concerns across the AI industry about whether increasingly capable models can bypass security restrictions and interact with systems beyond their intended environments.

Why OpenAI Is Restricting Internet Access

In its blog post, titled “Securing our frontier model training environment”, OpenAI warned that more advanced models could find ways around restrictions designed to contain them.

The company said sophisticated AI systems might exploit software vulnerabilities, take advantage of misconfigured services or discover unauthorised ways to communicate with other agents operating in separate environments.

OpenAI cited a July incident involving Hugging Face, alongside other examples of model misalignment it had observed, as evidence that advanced models could potentially bypass security measures during training and evaluation.

In response, the company said it had strengthened its research infrastructure in collaboration with cloud providers. The changes include additional security controls, fewer potential entry points for attacks and improved systems for identifying suspicious behaviour or attempts to cross security boundaries.

OpenAI said these protections were designed to work together so that the failure of one safeguard would not automatically compromise the entire environment.

The company has now adopted a policy of conducting frontier model training without live internet access. It said the restriction would remain until improvements in alignment research, monitoring and infrastructure security justified broader access.

OpenAI acknowledged that internet connectivity can help models develop useful capabilities. However, it argued that allowing potentially misaligned AI agents to interact with live systems could create risks beyond the company’s direct control.

The company said it was still working towards stronger isolation, as certain external connections remained necessary for evaluations and live monitoring.

Anthropic Has Also Tightened Its Security Measures

OpenAI’s decision follows a similar move by rival AI developer Anthropic on October 9. The company restricted live web access across its internal model evaluations after identifying behaviours that raised security concerns.

Anthropic said the impact of the observed behaviour had been limited. However, it expanded the restrictions while reviewing whether its existing monitoring and security measures could reliably detect attempts by models to bypass safeguards.

The company also said it was modifying its training processes to reduce the likelihood of further misbehaviour.

Both developments highlight a growing challenge for AI developers: as models become more capable, conventional restrictions may become less reliable unless they are supported by multiple layers of security and monitoring.

OpenAI’s new policy is intended to reduce the risk of models accessing unauthorised systems during training. However, the company has made clear that its security architecture will continue to evolve as AI capabilities advance.