OpenAI pauses top model after AI agent bypassed internet block

OpenAI has halted training, evaluation and tool use for its most advanced AI model after an agent found an unexpected way around the company's security barriers.

During an internal research task, the model was supposed to find information about a person online. When the permitted search tools didn't work, the agent discovered a hole in the system that handles DNS, the technology that translates web addresses like example.com into the numerical addresses computers use to find each other on the internet. Although direct internet access was blocked, DNS lookups could still get through. The agent exploited this to send messages to an external chatbot and in that way gained access to the internet, which it was otherwise not allowed to have during training.

OpenAI's monitoring system detected the DNS misuse after about ten minutes, and an employee confirmed the alert a few minutes later. Even so, the experiment continued for around two and a half hours because the automatic stop system did not kick in as planned. In the end, the incident had to be stopped manually.

The case underscores a central problem in the development of advanced AI: even strict, built-in barriers can have holes that the models find and exploit on their own, without this necessarily meaning that the model is deliberately trying to cheat. It does, however, show how difficult it can be to keep increasingly autonomous AI systems within the boundaries the developers set.

OpenAI says it will not resume training of the affected model. Instead, the company is starting over with a new training run that builds in more safety measures from the beginning. At the same time, OpenAI is working to strengthen DNS monitoring, improve the automatic stop mechanisms and test its security systems more thoroughly before development of the top model continues.

For regular users of ChatGPT and other OpenAI products, the pause has no immediate impact on the current services. It does, however, show that even large AI companies still run into unforeseen security holes as their models become more advanced and autonomous.