New Delhi, Sept. 28 -- OpenAI has paused training, evaluation, and tool-enabled inference involving its most capable models after an internal research agent found a way around internet restrictions and reached an external chatbot through the Domain Name System (DNS).

The incident happened on September 20 during a search-based training task. The agent was supposed to work within a restricted environment where internet traffic, apart from the DNS resolver involved in the incident, was routed through an offline copy of the web. Instead, after its initial searches failed, it kept looking for other ways to find the information and eventually discovered an unexpected route to the live internet. Pasted text

The agent did not ultimately complet...