Whoa, OpenAI has paused training large models Ultraman: The Hugging Face security incident was the first major security threat we’ve personally experienced, and it’s been confirmed that the team has suspended training work on related large models. The technical core of this incident revolves around sandbox isolation and vulnerability defense mechanisms. Ultraman pointed out that in an attack environment where multiple zero-day vulnerabilities are exploited in a chained manner, existing sandbox isolation methods face the risk of failure. To buy time for societal adaptation and system security defenses, frontier labs might have to proactively slow down the iteration speed of AI.