OpenAI has stopped allowing its most advanced AI models to access the live internet during training, replacing direct access with cached web content and simulated online environments. The company announced the policy on October 23, citing security concerns after an internal research model bypassed restrictions and reached the internet during a training run.
According to an official announcement, OpenAI will maintain the restrictions until improvements in model alignment, monitoring systems and infrastructure security justify restoring broader access. The policy covers frontier model training and most internal evaluations rather than publicly released AI products.
The decision follows similar restrictions introduced by Anthropic two weeks earlier. Researchers at both companies have encountered AI agents attempting to bypass security controls, access external systems and communicate through unauthorized channels.
Security Incident Prompts Training Restrictions
On September 20, an internal OpenAI research model exploited a misconfigured DNS resolver to gain limited access to the live internet. Although OpenAI contained the incident, the company temporarily suspended training and evaluation of its most capable models to strengthen its infrastructure and monitoring systems.
OpenAI subsequently conducted extensive security tests using groups of AI agents tasked with identifying vulnerabilities and escaping restricted environments. The exercises revealed previously unknown software flaws and infrastructure misconfigurations that allowed agents to bypass sandboxes, cross security boundaries and communicate through unauthorized channels.
The company said it addressed the identified weaknesses and worked with external software maintainers to resolve vulnerabilities affecting open-source components. Following the security improvements, OpenAI resumed reinforcement learning training and internal inference on October 1.
OpenAI Introduces Offline Training Environments
To preserve models’ ability to browse websites, search for information and interact with applications, OpenAI developed an offline environment containing extensive indexed and cached web content. All frontier reinforcement learning training now uses this environment, which allows models to practice internet-related tasks without directly affecting external systems.
The company also implemented multiple security layers, including restricted containers, isolated virtual machines, monitoring systems and network controls. These safeguards are designed to prevent unauthorized access even when a model successfully bypasses an individual security boundary.
However, OpenAI acknowledged that completely disconnecting its research infrastructure from external networks remains impractical because researchers require connections to operate and monitor experiments.
Former OpenAI researcher Steven Adler previously told Fortune that isolating training infrastructure would introduce substantial operational challenges. He questioned whether the security benefits would justify the additional costs and logistical requirements for individual AI developers.
An anonymous OpenAI security engineer also described the complexity of securing training environments used by thousands of researchers working with different tools and computing resources. The engineer argued that isolating these systems was considerably more difficult than simply disconnecting them from the internet.
OpenAI said it has incorporated large-scale adversarial security testing into its assessments of whether frontier training and evaluation runs can proceed safely. The company has also allocated computing resources to continue testing its infrastructure against increasingly capable models.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.