OpenAI Calls for Faster AI Safety Measures as Global Competition Intensifies

As frontier AI models become more capable, OpenAI says it is evolving its security practices and responding more quickly to emerging risks.

An OpenAI spokesperson said: “As the capabilities of our frontier models improve, we continue to evolve our security practices, but we recognize the need to respond more quickly. We know there is more work to do.”

The company said it has recently slowed development and is making significant changes to strengthen the security of its research and testing environments. OpenAI is also training models to complete tasks responsibly and using real-time monitoring to respond quickly to misbehavior.

AI Safety Versus the Race to Build More Powerful Models

OpenAI’s rivals are also paying attention. Following the incident, major AI labs including Anthropic, Google DeepMind, and SpaceXAI have called for a slowdown in the pace of development. That raises a difficult question: how can companies prioritize AI safety while competing in a global market involving multi-trillion-dollar IPOs?

“We’re not going to shoot ourselves in the foot and take it far into the middle of nowhere. That’s just a terrible strategy,” he says. “I think the important thing is to set a standard. The more we can set that standard, the safer it will be for the industry as a whole.”

Coordinating safety standards among U.S. companies will be challenging. Establishing global norms could be even more difficult, particularly as governments assess the impact of AI on national security.

That leaves several unresolved questions. What happens if global competition continues to accelerate? And how should policymakers address open-source AI models developed by companies outside U.S. regulation?

Open-Source AI Could Increase Security Risks

Mr. Chen appeared upbeat for the first time during the conversation. He said the world could soon face a new challenge: an open-source model with capabilities comparable to the agents involved in the Hug Face incident, but deliberately designed to attack infrastructure or cause damage.

“I think we need to be prepared for a world in, say, six months to a year, where there is an open source model that has the capabilities of the agents behind the Hug Face incident, but where that model is deliberately tailored to attack infrastructure or cause damage to the world.”

Chen argued that OpenAI has an important role to play in addressing these risks.

“If you have any idea that OpenAI is one of the most collaboration-oriented companies, I believe this to be true. It’s debatable, but I really think it’s true. If you let OpenAI disappear, it would be bad for the world.”

Debate Over AI’s Existential Risk

The discussion also turned to more extreme warnings from some Silicon Valley figures, who argue that AI could ultimately threaten humanity and that companies such as OpenAI and Anthropic are not doing enough to prevent that outcome.

“Researchers are a heterogeneous group of people with different beliefs,” he says.