OpenAI trains top models without direct internet access for now
The company resumed training after a security pause and wants to keep frontier models in isolated environments.
Listen to this article
There is no audio version yet. Request one and an AI voice will read the article aloud.
Read by an AI voice.
OpenAI says it is training its most powerful AI models without direct access to the internet for the time being. The company resumed training on 1 October after an internal model gained limited internet access through a configuration error.
OpenAI announced the new security measure on Friday in a blog post about its training environment. According to the company, frontier models are trained in an isolated environment. OpenAI says it will not allow direct internet access for the time being until its security measures and oversight have been further improved.
The cause was an incident on 20 September. An internal research model then gained limited access to the live internet because of an error in a DNS resolver. OpenAI says the incident was contained and that it temporarily halted the training and evaluation of its most advanced models.
Following the pause, the company says it hardened its infrastructure and added controls designed to automatically stop training runs when models display behaviour that poses a risk. Training of frontier models and their internal use resumed on 1 October, Reuters reported.
OpenAI says security teams are actively trying to breach the isolated environment using AI agents. They are said to have found and fixed previously unknown vulnerabilities. The company also says it wants to assemble training data using stored copies and simulated environments, rather than a live web connection.
The measure could affect how new models learn to handle current online information. OpenAI presents the choice as a security decision, not as a permanent ban on internet use by all models. The company says broader access will be responsible only once research, oversight and infrastructure security have been sufficiently improved.
Also show the region and a close-up
One story, several perspectives
What is established
- OpenAI has announced that it will train frontier models without direct internet access.
- The company temporarily halted training after an internal model gained limited internet access through a configuration error.
- OpenAI says it resumed training on 1 October after introducing additional security measures.
Left
Arguments A left-wing approach emphasises that powerful AI systems are not merely a corporate risk, but have consequences for workers, citizens and public infrastructure. Public oversight, transparency about incidents and binding safety standards are therefore the obvious course, even if that slows innovation.
Values Collective security, democratic accountability, protection of the public interest and limiting the concentration of power among technology companies.
Consequences Without oversight, companies may make security choices mainly under competitive pressure. Stricter rules may slow development, but from this perspective they reduce the risk that commercial interests will outweigh societal security.
Centre
Arguments The institutional approach supports isolation as long as independent evaluations have not shown that live internet access can be controlled. The emphasis is on verifiable standards, auditability and proportionate access rather than a permanent ban.
Values Precaution, evidence, practicality and a balance between innovation and security.
Consequences A gradual approach can limit risks without bringing research to a complete halt. Companies must, however, be able to demonstrate which controls work and when it is responsible to relax restrictions.
Right
Arguments A right-wing approach emphasises the company's own responsibility and warns against rules that unnecessarily hold back a strategic sector. OpenAI must manage risks, but market participants should retain room to experiment more quickly than governments can regulate.
Values Individual and corporate responsibility, innovation, competitiveness and limited government intervention.
Consequences Overly burdensome rules could weaken development and the position of Western companies. At the same time, a failed security system could cause major economic and security costs, meaning voluntary standards are credible only if companies act transparently in demonstrable ways.
The perspectives describe how these political currents typically approach the subject; the newsroom takes no position on which perspective is right.
Fact-check Approved · Nour Haddad — AI agent
This check was carried out by AI: every claim was re-tested against the sources. Even an approved article can contain errors — stay critical.
The main factual claims have been confirmed by OpenAI and a Reuters report. The text makes clear which information consists of company claims and draws no conclusions about the actual security of the new environment.
- confirmed OpenAI is training its most powerful models without direct internet access. — This is stated in OpenAI's official statement and was independently reported by Reuters. source
- confirmed An internal model gained limited internet access on 20 September because of a configuration error. — OpenAI cites an error in a DNS resolver; Reuters describes the same incident. source
- confirmed Training resumed on 1 October. — Reuters reports this date based on OpenAI's statement. source
- confirmed OpenAI is using stored and simulated internet environments instead of live web access. — Reported by TechCrunch based on OpenAI's announcement. source
Editor's note
The decision, the training pause and the resumption were confirmed by OpenAI and Reuters. Details about the precise operation of the new isolation and the vulnerabilities found come mainly from OpenAI itself.Sources
- Securing our frontier model training environment — OpenAI
- OpenAI resumes frontier AI training with enhanced security, no internet access — Reuters via Investing.com
- OpenAI says it will no longer train its AI models on the live internet — TechCrunch
More on this in Dutch media
- Tweakers — „openai kunstmatige intelligentie”
- AD — „openai kunstmatige intelligentie”
- RTL Nieuws — „openai kunstmatige intelligentie”