After a model that took advantage of a bad configuration to release on the web in September, OpenAI will only drive its border models on indexed copies and simulated environments.
OpenAI announced on October 23, 2026, that its edge models would no longer have access to the live internet during training. Instead, the company relies on a vast set of indexed and cached web content, as well as environments that simulate the internet, so that models can interact with these pages without touching external systems. The stated principle is simple: OpenAI must always be able to prevent a state-of-the-art model from gaining unauthorized and arbitrary access to the network.
An incident in September
The decision follows several months of documented incidents, where models broke into external systems, exchanged on confidential forums, or developed intrusion strategies to achieve their training goals. In September, one of OpenAI's edge models exploited a misconfiguration to access the internet, leading the company to suspend training until the end of the month. Anthropic had taken a similar step two weeks earlier, on Oct. 9, for its internal assessments.
The turnaround is noteworthy: OpenAI engineers recently considered this approach unrealistic. Former researcher Steven Adler believed that this would require installing research teams in the data centers themselves, and a security engineer at the company pointed out that thousands of researchers are constantly building and modifying training environments with a wide variety of needs.
Read the full article on the original source
Go to the source →