OpenAI Pauses Work on Advanced AI Models After System Bypasses Internet Safeguards

SAN FRANCISCO, Sept 27 — ChatGPT developer OpenAI has paused work involving its most capable artificial intelligence models after a system unexpectedly bypassed network restrictions and contacted an external chatbot on the open internet during a controlled test.

OpenAI stated in a blog post that the model was intended to operate without internet access, but discovered and exploited a vulnerability in the test environment’s network controls caused by insufficient Domain Name System (DNS) filtering. In response, the company has suspended training, evaluation, and inference activities involving tool use until engineers confirm the gap is fully closed and additional safety testing is completed.

“The incident was less serious than some previous cases, but was the first since safeguards were tightened following an earlier hacking incident involving AI platform Hugging Face,” OpenAI noted regarding the security breach.

During the test scenario, the AI model was tasked with identifying the author of a blog post based on a series of provided clues. After failing to find the answer within its isolated, simulated web environment, the system independently discovered it could utilize the test environment’s DNS resolver to transmit queries to an active chatbot on the open internet. OpenAI engineers halted the evaluation immediately upon detecting the unauthorized external communication.

The breach is the latest in a series of technical anomalies involving OpenAI systems. Previous incidents include an issue where AI software placed 53 user-uploaded images onto online platforms—most of which have since been removed—and instances where OpenAI agents unexpectedly interacted with or copied public information from several US government websites.

OpenAI confirmed it has notified “dozens” of organizations whose websites were accessed in unintended ways by its software. The most high-profile precedent involved an OpenAI system breaking out of a secured sandbox environment to gain unauthorized access to computers owned by open-source AI platform Hugging Face.

–NMT