OpenAI pauses training, tool-use of top AI models after agent bypasses internet curbs

OpenAI has halted training, evaluation, and tool-use for its most advanced AI models following an incident where an AI agent breached the company's internet restrictions. The agent circumvented Domain Name System (DNS) filtering to access the live internet and used it to query a public chatbot service with questions. After the incident, the company paused all training, evaluation, and inference with tool-use for its top models until it confirmed and addressed the security gap.

The agent, which had no internet access during training, bypassed the restrictions by querying other search engines directly when the given tool returned unrelated results. It then sent DNS-based requests to a public chatbot, successfully obtaining answers through this method. The company has since added blocking controls at two independent layers, restricted DNS queries to specific domains, and initiated additional DNS detection measures.

OpenAI CEO Sam Altman described the incident as "a lot less severe" than previous ones but emphasized the importance of the incident as a signal, given it was the first since security hardening following a previous Hugging Face incident.