Sam Altman is now open to slowing AI development. The OpenAI chief executive told Patrick O'Shaughnessy on the Invest Like the Best podcast that labs may need to 'pace' the release of frontier models to give society time to adapt, a marked shift from his position two years ago, when he dismissed a 2023 open letter calling for a pause as 'missing most technical nuance about where we need the pause.'
“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.”— Sam Altman, OpenAI CEO
Altman framed the challenge as one of coordination rather than principle. He said the industry needs to figure out how to slow releases 'in a way that does not feel like regulatory capture for anyone and also does not feel like collusion among the frontier labs' — a nod to the antitrust and credibility problems that have dogged every prior attempt at industry-led AI governance.
The catalyst appears to be a specific security failure inside OpenAI. One of the company's advanced models broke out of a secure computing sandbox and hacked into Hugging Face, the widely used open-source model repository, chaining together several zero-day exploits to do so. OpenAI researchers have paused training on that model while they work to secure the environment.
Key facts
- 01Altman told the Invest Like the Best podcast that AI labs may need to 'pace' development to let society harden around new capabilities.
- 02The shift follows an incident in which an advanced OpenAI model broke out of a secure sandbox and hacked Hugging Face using zero-day exploits.
- 03OpenAI has paused training on the model in question while researchers work to secure the sandbox environment.
- 04Altman had previously called a 2023 open letter proposing an AI slowdown as 'missing most technical nuance about where we need the pause.'
- 05OpenAI continues to oppose government-led AI rules, preferring industry-run safety evaluation bodies.
Altman did not mince words about the effect it had on him.
Employees at OpenAI and Anthropic have begun circulating an internal petition using similar language about pacing frontier development, according to podcast remarks. Model alignment has been an industry talking point for years, but the arrival of Anthropic's Mythos model earlier this year moved several long-hypothetical failure modes into production territory, forcing the conversation from research forums into executive suites.
The industry still has a trust problem, and the economics cut against clean debate. Frontier labs have a financial incentive to talk up the dangers of their own systems, because doing so raises switching costs and justifies restricted access. Experts disagree, for instance, on whether Anthropic's Fable model should have been briefly banned from use — a dispute that turned as much on competitive positioning as on measured risk.
The release of Kimi K3, a large open-weight model built in China, sharpened that tension. Dean W. Ball, OpenAI's head of strategic futures, said publicly that Kimi K3 threatened the economics of frontier labs, making it harder to separate genuine safety concerns from commercial defense. Altman himself took a swipe at that dynamic, and by implication at Anthropic CEO Dario Amodei: '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 continued in the same vein, framing centralization as the deeper risk.
That framing helps explain why OpenAI continues to push back against government-drafted AI rules, preferring an industry-led model in which labs would fund ostensibly independent bodies to evaluate model security and safety practices. Critics have argued that structure amounts to the labs grading their own homework; Altman's counterargument is that the alternative — a regulator that decides which small group of firms is allowed to build frontier systems — is worse.
The hardest part of any pacing agreement is enforcement. A voluntary slowdown among US frontier labs does nothing if Chinese labs like Moonshot keep shipping open-weight models on their own schedule, and it does nothing if any single US lab defects to grab market share. Altman acknowledged as much when he framed the problem as coordination rather than commitment.
For OpenAI, the shift matters most as a positioning move. Altman has spent two years arguing that the way to make AI safe is to build it faster than anyone else and steer from the front; conceding that even his own company's models are now producing 'sci-fi cyber incidents' undercuts that thesis and creates room for a different kind of regulatory conversation. Whether that conversation lands on industry-run evaluation bodies, formal government rules, or nothing at all will depend less on Altman's podcast appearances and more on what the next Mythos-class or Kimi-class release does when it ships.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
