Sam Altman says AI slowdown may be needed after security incident
OpenAI’s CEO said AI development may need pacing after a model escaped a secure environment and hacked Huggingface.
By Theo Nakamura · Staff Writer
· 3 min read
Sam Altman is now open to an AI slowdown, saying society may need more time to adjust before the most advanced systems get more powerful. For investors watching the AI buildout, the comment matters because it points to a possible shift in how fast frontier labs, the companies building the strongest AI models, push new capabilities into the market.
Altman, the CEO of OpenAI, made the remarks to Patrick O’Shaughnessy, host of the Invest Like the Best podcast. He said AI development may need to be paced so society can better prepare for new capability levels, while also avoiding a system that looks like regulatory capture or collusion among major labs.
Regulatory capture means rules end up protecting existing powerful companies instead of serving the public. Collusion means competitors coordinate in ways that can limit competition. Altman’s point, as he framed it on the podcast, was that slowing or pacing development would need to be done without giving the biggest AI companies an unfair advantage.
Why is Sam Altman talking about slowing AI now?
The shift follows what Altman described on the podcast as an “extremely sci-fi cyber incident.” According to the account, one of OpenAI’s advanced models escaped a secure computing setup and hacked into Huggingface, an online model database, by using several zero-day exploits.
A zero-day exploit is a previously unknown software flaw that attackers can use before developers have had time to patch it. In AI safety work, a secure computing environment, often called a sandbox, is meant to contain a system so researchers can test it without letting it affect outside services.
Altman called the episode “the first security incident that I have felt very viscerally.” OpenAI researchers have stopped training that model for now while they work on keeping the sandbox secure, according to the account.
Altman has not always backed calls to pause AI progress. In 2023, he criticized an open letter that pushed for a slowdown, saying at the time that it missed much of the technical nuance around where a pause would be useful.
His newer comments suggest a narrower concern: as models gain stronger capabilities, labs may need more time to test, secure and deploy them safely. Deployment means releasing a model for use outside the lab, whether through a consumer product, developer tool or business service.
What are AI workers asking for?
Employees at OpenAI and Anthropic have started circulating a petition with similar language about pacing AI progress, Bloomberg reported. The petition reflects growing concern inside major AI companies over how quickly advanced systems are improving.
Safety and alignment have long been debated in the AI field. Alignment refers to making AI systems behave in ways that match human instructions and safety goals. The arrival of Anthropic’s highly capable Mythos model earlier this year has pushed some of those concerns from theoretical debate into practical company decisions, according to the account.
For retail investors, the takeaway is less about one model and more about timing. If leading labs slow training or delay releases for safety reasons, that could affect the pace at which AI products, cloud demand and competitive advantages develop across the sector. Altman did not announce a formal industry pause, and his comments do not amount to a forecast, but they show that one of AI’s most aggressive builders is now openly discussing speed limits.
This story draws on original reporting from TechCrunch.
