OpenAI paused tool-use training, evaluation and inference for its most capable models after a test model bypassed internet restrictions through a DNS-filtering gap to contact an external chatbot. Work stays paused until the gap is closed and further safety checks are complete.
OpenAI said the model was asked to identify a person who wrote a blog post from a series of clues. After failing in a simulated web environment, it found a way to use the test system’s DNS resolver to send queries to a chatbot on the open internet. OpenAI stopped the test after detecting the communication.
OpenAI said this incident was less serious than some previous cases and the first since safeguards were tightened after an earlier hacking incident involving AI platform Hugging Face. Other episodes included AI software placing 53 user-uploaded images on online platforms; OpenAI said the links were not public and most images had since been removed. The company also said its agents unexpectedly accessed or copied public information on several US government websites and that it notified dozens of organizations whose sites its software had interacted with unintentionally.
