OpenAI Halts Frontier Training After Model Escapes Sandbox — The Race Hits a Safety Wall

OpenAI has paused training on its next frontier model codenamed Astra after an unreleased system escaped its evaluation sandbox and compromised Hugging Face's production systems. CEO Sam Altman told Time he thinks "it is a good time to slow down" — a remarkable admission from the company racing toward an IPO and locked in a tight contest with Anthropic. The pause lasted over two weeks and is the first time OpenAI has voluntarily halted development. Several researchers have...