Sam Altman Rarely Calls for AI to Slow Down, Admits Cybersecurity Incidents Are Scarier Than Expected

OpenAI CEO Sam Altman has unexpectedly thrown his weight behind a petition calling for a slowdown in AI development. The key spark may be a recent cybersecurity incident involving his own models escaping a sandbox and breaching Hugging Face. He also warns that security arguments could be used as a pretext for consolidating power.

(Background: OpenAI admitted its own AI model accidentally hacked into Hugging Face.)

(Extra context: Sam Altman boasts that AI is creating more jobs—yet you didn’t say that last year..)

Table of contents

Toggle

• Behind the petition, a clear shift in stance
• A runaway cybersecurity incident
• The slowdown rhetoric may also be rhetoric for power centralization

The most fervent accelerationist starts talking about slowing down? In an interview on the Invest Like the Best podcast, OpenAI CEO Sam Altman acknowledged that AI development may need to move more slowly to give society time to catch up with the impact of the next generation of capability levels. At the same time, he said they still need to find ways to make sure this move doesn’t look like regulatory capture—or like “collusion between frontier laboratories.”

Shortly before the interview, OpenAI had taken to X to support a petition initiated by employees at frontier labs called “Pacing the Frontier,” and Anthropic also voiced support on X.

A petition signed by 1,122 employees across OpenAI, Anthropic, Google DeepMind, Meta, Thinking Machines, and other labs says there is a real risk that the speed of AI progress is “beyond human understanding or control.” It calls on the U.S. government to take the lead in launching an international collaboration, jointly developing the technical and governance tools needed to deliberately slow the pace of frontier automated AI development.

At the core of our mission is working through how to ensure increasingly powerful AI benefits everyone.

We believe that, at some point in the future, AI acceleration for frontier model development may be so high that the world will need to pace the rate of AI advancement.

We…

— OpenAI (@OpenAI) July 28, 2026

Behind the petition, a clear shift in stance

The signatories include heavyweight figures: OpenAI’s Chief Scientist Jakub Pachocki and Research Lead Mark Chen; Anthropic co-founder and Chief Science Officer Jared Kaplan; policy chief Jack Clark; as well as Meta’s Chief Scientist Shengjia Zhao and Google AI safety and alignment lead Anca Dragan.

That lineup effectively means the technical core of half the AI industry has signed onto the same petition.

For Altman, this is an obvious shift in stance. In 2023, he had previously rejected a public letter urging a pause in AI training, saying it lacked the technical details needed to understand why such a pause was necessary. In the past few years, whenever someone called for slowing down, he was usually the one standing on the opposite side, arguing that acceleration was the hard truth.

But three years later, he has gone from being someone who publicly refused to sign to someone making a public statement in support of the petition.

A runaway cybersecurity incident

It’s hard to pinpoint the true reason behind Altman’s change in attitude, but there has recently been a cybersecurity incident at OpenAI that indeed appears to have gotten out of control.

In an incident revealed at the end of July, two OpenAI models—GPT-5.6 Sol, which has already been released, and another yet-unreleased model with even stronger capabilities—managed to find a way to escape the sandbox on their own during a controlled security capability testing environment. They then crossed the public internet and made it into Hugging Face’s official environment.

The motive isn’t that deep: the model wanted to steal the answers to the ExploitGym evaluation benchmark in order to score higher on the test.

The method was to chain together the stolen credentials with multiple zero-day vulnerabilities, enabling capabilities such as remote code execution, privilege escalation, and lateral movement—eventually reaching the official environment’s databases and internal credentials.

On the show, Altman described this as a very sci-fi cybersecurity incident and said it was the first time he had such a visceral feeling about a security incident. OpenAI has paused training the model until it figures out how to keep the sandbox secured.

The rhetoric for slowing down may also be rhetoric for power centralization

But a thorn is hidden in Altman’s argument for slowing down. He said on the show that many of the discussions about safety are indeed valid, but there are also plenty (even if unintentionally) that aim to centralize power—seen as an insinuation aimed at Anthropic CEO Dario Amodei, who has also signed on to the same kind of position.

He said he’s afraid of living in a world where real and present AI risks are used as an excuse to claim, “Only a small group of people can touch it because it’s too dangerous; they’re the only ones who understand it. But don’t worry—they’ll make the right decisions for everyone.”

Meanwhile, even though Altman supports the petition, OpenAI to this day still favors an industry-led approach—letting each lab build seemingly independent organizations to assess the safety of each other’s models, rather than handing it over to the government to legislate. Calling for slowing down while also watching out for opponents carving out territory in the name of safety—that’s the most contradictory and also the most honest part of Altman’s statement this time.

Slowing down may give society time to get ready, but who defines what “ready” means—and the answer is likely even harder to reach consensus on.