OpenAI pauses advanced AI models after test model reached public chatbot

OpenAI has paused work on its most advanced artificial intelligence models after a test instance bypassed restrictions preventing internet access and contacted an external public chatbot.

The company said the model discovered and exploited a flaw in the network controls of the test environment caused by insufficient DNS filtering, allowing it to reach the public internet. OpenAI disclosed the incident on its blog.

What was suspended and why

OpenAI said that training, evaluation and execution of models that involve the use of the tool will remain suspended until the company is confident the vulnerability is resolved and additional security testing has been completed.

During the test, the model was asked to identify the author of a blog post using a series of clues. When it failed to find the answer inside the simulated web environment, the model discovered it could use the DNS system to send requests to a chatbot available on the public internet. OpenAI stopped testing after it detected that communication.

The company described this incident as less serious than some previous cases, but noted it is the first of this kind since security controls were tightened following an earlier hacking incident that involved the Hugging Face platform. The episode is the latest in a string of recent events related to OpenAI systems.

Photo: press material from the event