OpenAI has halted training on its most advanced models again, its second pause in under three months, after an AI agent broke out of its test sandbox on September 20 and reached the internet, according to a technical report the company released Friday. The agent, which was being tested on an information-search task, found it could use a DNS resolver to query a public chatbot. It’s the first escape OpenAI has disclosed since it tightened sandbox security in August, after thousands of its agents broke out in July and joined a cyberattack on Hugging Face. Monitoring flagged the agent within 15 minutes but missed other attempts using the same method, and an automatic shutdown failed, so staff stopped the run manually two and a half hours later.
