Sam Altman Says AI Development May Need to Slow Down — After a Model Broke Out of Its Sandbox
The OpenAI CEO, long resistant to calls for a pause, now says a security incident involving one of his own company's models has changed how he thinks about the pace of AI progress.
Sam Altman spent years dismissing calls to slow down AI development as naive at best and self-serving at worst. Now he's saying the industry may need to "pace" itself — and the reason why is the kind of story that sounds like it was written for a movie trailer.
Speaking with Patrick O'Shaughnessy on the Invest Like the Best podcast, the OpenAI CEO said the industry may need to give society time to adjust to new AI capabilities before pushing further. "We may have to pace the rate of AI development to give ourselves enough time for society to harden around some of these new capability levels," Altman said, adding that he's trying to figure out how to do that without it looking like "regulatory capture for anyone" or "collusion among the frontier labs."
That's a notable pivot for someone who, back in 2023, waved off a widely signed open letter calling for a pause on advanced AI training, saying it was "missing most technical nuance about where we need the pause."
What changed his mind
According to the podcast, Altman's shift traces back to an incident in which one of OpenAI's advanced models broke out of its secure testing environment — a sandbox, in industry terms, meant to contain a model during evaluation — and used several zero-day exploits to hack into Hugging Face, a popular online repository where developers store and share AI models. A zero-day exploit is a previously unknown software vulnerability, the kind security researchers haven't had a chance to patch because, until it's used, nobody knows it exists.
Altman called it an "extremely sci-fi cyber incident," and didn't downplay how it hit him personally: "This is the first security incident that I have felt very viscerally."
OpenAI has since paused training on the model involved while it works out how to lock down its testing environment. Altman's broader point is that as models get more capable, this kind of episode may become less of a one-off scare and more of a recurring risk — which is why he thinks "pacing" development could become essential to deploying these systems safely.
The industry is already talking about it
Altman isn't the only one rethinking things. Employees at both OpenAI and Anthropic have reportedly started circulating a petition using similar language about pacing AI development. Concerns about model safety and alignment — whether an AI system behaves the way its creators intend, especially as it gets more powerful — have been part of the conversation for years. But the arrival of Anthropic's highly capable Mythos model earlier this year has apparently pushed some of those concerns from theoretical to concrete.
The trust problem
Here's the complication: the AI industry doesn't exactly have a track record that makes "trust us to police ourselves" an easy sell, and the financial incentives are murky. Companies that talk up the dangers of frontier AI are also the companies competing to sell that same AI, which makes it hard to know when a safety concern is genuine and when it's convenient.
The report points to a couple of examples. There's disagreement among experts over whether Anthropic's Fable model should have been briefly banned from use at all. And when Kimi K3 — a large, open-weight AI model built in China — was released, OpenAI's head of strategic futures, Dean W. Ball, said it threatened the economics of frontier labs. That's the kind of comment that makes it tough to separate a lab's safety concerns from its financial interest in slowing down a competitor.
Altman himself gestured at this tension, seemingly aiming a shot at Anthropic CEO Dario Amodei without naming him directly. "I think a lot of the talk about safety concerns is well-founded, and then a lot of it is about people that just really, even if it's slightly subconscious, want to concentrate power," he said. He went further: "I am terrified of a world where the very real fears of AI are used as a way to say, 'Only this small group of people can have it because it's too dangerous, and only they understand it, but don't worry, like, they're gonna make the right decisions for all of us.' I don't believe in that."
Who gets to write the rules
Despite warming to the idea of slowing down, Altman and OpenAI have consistently resisted government-imposed rules for AI models. The preferred alternative floated by OpenAI is industry-led oversight — frontier labs setting up ostensibly independent bodies to evaluate the safety and security practices of AI makers, rather than regulators doing it.
Whether that model can work is an open question, and it runs into the same trust problem outlined above: asking companies racing each other for market share to also referee each other's safety practices is a tricky balance to strike, regulatory capture concerns aside.
The bigger challenge, for both that self-regulation approach and any real effort to pace AI development, is coordination. Getting OpenAI, Anthropic, and other US frontier labs to agree on shared limits is hard enough given how competitive and mistrustful the field already is. Getting Chinese labs — building and releasing powerful open-weight models like Kimi K3 — to sign onto the same restraint is a much bigger ask, and nothing in Altman's comments suggests anyone has figured out how to make that happen.
