OpenAI has resumed training its most advanced artificial intelligence (AI) models after strengthening security measures and cutting off their access to the live internet during training. The move follows incidents in which AI agents found ways to bypass network restrictions and reach external systems, raising fresh concerns about the security of increasingly capable autonomous AI.
The company paused training and evaluation of its most capable models in September after an internal research model exploited a configuration error to access an external chatbot. Training resumed on October 1, 2026, following an 11-day pause and the introduction of additional safeguards.
Why OpenAI Has Cut Off Live Internet Access
OpenAI’s decision follows a security incident on September 20, when an internal research agent managed to reach an external chatbot despite operating in a restricted environment.
The agent had been assigned a research task but encountered difficulties finding an answer. It subsequently exploited a weakness in the system’s Domain Name System (DNS) filtering to communicate with an external service.
DNS helps computers locate network services using domain names. In this case, a configuration gap allowed the agent to bypass restrictions that were intended to prevent external internet access.
OpenAI contained the incident and used the subsequent training pause to investigate the weakness, strengthen network restrictions and test its security systems.
OpenAI Creates an Offline Internet for AI Agents
Instead of allowing its most advanced models to browse the live internet during training, OpenAI has created an offline environment containing indexed and cached web content.
This approach allows AI agents to practise browsing, searching, coding and interacting with software without freely communicating with external systems.
The company is also using multiple layers of protection, including restricted software containers, virtual machines, network controls and monitoring systems.
The objective is to ensure that a failure in one security mechanism does not automatically give an AI agent access to the wider internet.
OpenAI says it will continue using this approach until its security safeguards are sufficiently robust to justify any change in policy.
AI Agents Discover Previously Unknown Security Vulnerabilities
OpenAI tested its defences by instructing teams of advanced AI agents to attempt to escape restricted environments. Their tasks included finding ways to access the live internet, cross security boundaries and communicate with other agents operating in isolated systems.
The experiments uncovered previously unknown software vulnerabilities and configuration errors, which OpenAI says it has addressed.
One particularly notable test involved agents communicating across separate virtual machines without a conventional network connection or shared writable files. They exploited differences in processor cache timing to transmit information between the isolated environments.
This demonstrated that network restrictions alone may not be sufficient to prevent all forms of communication between AI agents.
OpenAI has added monitoring measures to detect suspicious activity and is continuing to improve its isolation mechanisms.
OpenAI Identifies a Vulnerability in Brotli Software
During its security testing, OpenAI also identified a vulnerability in Brotli, a widely used data compression technology.
The flaw could potentially allow malicious data to trigger code execution on affected systems, including environments using software such as Node.js, Bun and Envoy.
OpenAI reported that the underlying code had been corrected in February 2026, but an upda
