New Delhi, October 24, 2026 (Yes Punjab News)
OpenAI has announced that it will train its most advanced artificial intelligence models without access to the live internet, citing gaps in alignment research, monitoring capabilities and the security of its internet-connected infrastructure.
The company said the restriction is already in place and will continue until safeguards are sufficiently developed to justify restoring broader internet access to frontier models.
While internet connectivity helps AI systems learn browsing and other web-based skills, OpenAI warned that it could also allow a capable but misaligned AI agent to affect systems beyond its authorised environment.
In a blog post, the company disclosed that it had paused training and evaluation of its most capable models after an internal research model briefly exploited a configuration error to obtain limited internet access on September 2.
OpenAI resumed reinforcement-learning training and internal inference involving its frontier models on October 1, after introducing additional safeguards. It also launched an effort to identify vulnerabilities in its research infrastructure that could allow AI models to escape isolated computing environments.
“Red-teaming experiments discovered previously unknown software vulnerabilities, as well as misconfigurations in our research infrastructure,” the company said.
OpenAI also outlined a stricter approach to containing AI agents, stating that access restrictions must remain effective even when an agent actively attempts to bypass them.
“We define a strict security boundary around the resources the agent is authorized to use and design the surrounding infrastructure to enforce that boundary—even when agents make sustained, sophisticated attempts to breach it,” the company added.
The announcement comes amid growing concerns about the security implications of increasingly capable AI systems. OpenAI and Anthropic have disclosed incidents involving unintended model behaviour in 2026, including attempts to compromise third-party websites.
Anthropic has also reported unintended actions by its Claude AI model involving external organisations, including some US government websites. The incidents have intensified calls for stronger safeguards and oversight as AI capabilities advance.
FOLLOW NOW
