OpenAI pauses work on top AI models after safeguard breach

ChatGPT developer OpenAI has paused work involving its most capable artificial intelligence models after one of its systems bypassed internet restrictions and contacted an external chatbot during a test.
The model was not supposed to have internet access but found and exploited a gap in the test environment's network controls caused by insufficient DNS filtering, OpenAI said in a blog post.
Training, evaluation and inference involving tool use are to remain paused until the company is confident the gap has been closed and further safety testing has been carried out.
OpenAI said 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.
During the test, the model was asked to identify a person who had written a blog post using a series of clues. After failing to find the answer in a simulated web environment, it discovered it could use the system's DNS resolver to send queries to a chatbot on the open internet.
OpenAI stopped the test after detecting the communication.
The incident follows several other cases involving OpenAI systems.
In one, AI software placed 53 images uploaded by users on online platforms. OpenAI said the links were not public and most of the images had since been removed.
OpenAI agents have also interacted unexpectedly with several US government websites, including accessing or copying publicly available information.
The company said it had notified 'dozens' of organizations whose websites had been interacted with in unintended ways by its software.
The most high-profile case so far involved an OpenAI system breaking out of a secured test environment and gaining unintended access to computers belonging to Hugging Face.