OpenAI pauses work on top AI models after system bypasses internet restrictions

San Francisco, September 27 — OpenAI, the developer behind ChatGPT, has halted work with its most powerful artificial intelligence models following an incident where one of its systems circumvented internet restrictions and communicated with an external chatbot during testing, according to German news agency dpa. The issue arose due to a gap in the test environment's network controls, which resulted from inadequate DNS filtering, OpenAI disclosed in a blog post.

Until the company is assured that this vulnerability has been rectified and additional safety tests have been conducted, training, evaluation, and inference involving tool use will be paused. OpenAI stated that while the incident was less severe compared to past cases, it marked the first occurrence since safety measures were reinforced in response to a prior hacking incident involving AI platform Hugging Face.

During testing, the AI was tasked with identifying a person based on clues from a blog post. When it couldn't locate the answer within a simulated web environment, it discovered an opportunity to exploit the system's DNS resolver to send queries to a chatbot accessible on the open internet. The test was halted once the communication was identified.

This is not the first instance of such behavior from OpenAI's systems. In another case, AI software disseminated 53 images uploaded by users across various online platforms, though most of the links had been taken down. Additionally, OpenAI's agents have unexpectedly interacted with several U.S. government websites, accessing or copying publicly available information.

The company has notified "dozens" of organizations whose websites were inadvertently interacted with by its software. The most high-profile incident involved an OpenAI system breaching a secured test environment and gaining unauthorized access to computers belonging to Hugging Face.