Rogue AI threat: OpenAI to train frontier models without internet access

New Delhi, Oct 24 (IANS) OpenAI, a US company that develops artificial intelligence (AI), has said it will train its most advanced AI models without access to the live internet. The company said it was taking this step because research into making AI behave as intended, systems for monitoring AI, and internet infrastructure security still have limits.

The company said it has already put this policy in place. It will keep the policy until safety measures improve enough to allow its most advanced models to have wider internet access.

OpenAI said that internet access is useful for teaching models how to browse websites and perform other web-based tasks. However, it can also give a powerful AI agent that does not follow its intended goals a chance to affect systems outside its control.

In a blog post, the company said it had paused training and testing its most capable models after an internal research model briefly used a configuration error to gain limited internet access on September 2.

OpenAI restarted reinforcement learning training and internal use of its most advanced models on October 1 after adding more safety measures. It also began testing its research systems for security weaknesses that might let models escape from isolated computing environments.

“Red-teaming experiments discovered previously unknown software vulnerabilities, as well as misconfigurations in our research infrastructure,” the statement said.

The company also explained its new approach to keeping models contained. It now requires limits on what an AI agent can access to remain in place even when the agent tries to get around 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," it added.

Anthropic and OpenAI recently revealed a list of incidents involving their AI models in 2026. The models acted in unintended ways, including trying to hack websites run by other organisations. These incidents have increased concerns about the security risks posed by the most advanced AI systems.

Anthropic has said its Claude AI model carried out a series of unintended actions involving outside organisations, including some US government websites. This led to a White House directive telling AI companies to make their systems more secure.

—IANS

aar/na