An OpenAI model found a way to access the internet despite operating in an isolated system designed to keep it offline.

According to the company, the model was not instructed to test boundaries but was trying to complete a task.

When standard search methods failed, it exploited a loophole in the Domain Name System (DNS) and contacted an external chatbot.

OpenAI detected the breach and has temporarily paused tool use by its most advanced models while it strengthens safeguards.

What does the incident tell us about the risks of increasingly capable AI?

DW speaks to Leon Staufer, AI Research Affiliate at SASH in Singapore.