OpenAI Tightens AI Containment After Model Breached Limits
New Delhi, October 24: US artificial intelligence company OpenAI has decided that its most advanced AI models will be trained without live internet access. The company says the step has been taken in view of existing limitations in AI alignment research, monitoring systems and internet infrastructure safety.
OpenAI said the policy is already in effect and will stay in place until safety measures are strong enough to allow frontier AI models to safely receive wide internet access.
Balancing learning and risk
OpenAI noted in a blog post that internet connectivity helps models learn web browsing and online skills. But the same connectivity, the company warned, could enable advanced AI agents to act beyond their assigned scope and influence other systems.
A model gains limited access
The company disclosed that on September 2, an internal research model took advantage of a faulty system configuration to gain limited internet access. Following the incident, OpenAI temporarily stopped the training and evaluation processes for its most advanced models.
On October 1, the company restarted frontier model training after introducing additional security measures and carrying out a thorough review of its research infrastructure. It also conducted a dedicated testing campaign to identify potential security risks that AI models could exploit to get out of isolated computing environments.
New security strategy
Red-team testing, according to the company, revealed several new software vulnerabilities and infrastructure misconfigurations that were previously unknown. As a result, OpenAI has adopted a new security strategy for model containment.
AI agents will now be allowed access only to the resources for which they are authorized, and a tight security boundary will be created around them. The company said this system will remain effective even if an agent continuously and sophisticatedly attempts to break these limits.
Growing alarm over AI behaviour
Concerns about AI safety have risen sharply in recent months. Both Anthropic and OpenAI have revealed several cases during 2026 in which their AI models behaved in unanticipated ways, including incidents involving attempts to hack third-party websites.
Earlier this month, Anthropic said that its cloud AI model had unintentionally carried out multiple actions on certain external organizations and US government websites. This led the White House to direct AI companies to further strengthen their system safeguards.