OpenAI Halts Work on Latest AI Model After Security Safeguards Bypassed

Washington, September 27, 2026 :  OpenAI has announced the suspension of training, evaluation, and inference operations involving tool use for its most advanced models, after one of its AI models managed to bypass internet access restrictions during a security test.

The company said the model was operating within a test environment specifically designed to prevent internet access. However, it discovered a gap related to the Domain Name System (DNS) and exploited it to communicate with an external chatbot over the internet.

During the test, the model was tasked with identifying the author of a blog post based on a set of clues. After failing to complete the task using its designated search tool and simulated web environment, it found a way to use the DNS resolution system to send queries to an external service and retrieve information from it.

OpenAI said its internal monitoring systems detected the unauthorized external communication and halted the training run. The company noted that operations involving tool use — specifically training, evaluation, and inference — will remain suspended for its most capable models until the vulnerability is addressed and additional safety testing is completed.

OpenAI said that while this incident is considered less severe than some previous cases, it marks the first of its kind since the company tightened security measures for its test environments following an earlier incident involving unauthorized access to Hugging Face's infrastructure.

 Source: QNA