OpenAI chief executive Sam Altman has criticised the way rival artificial intelligence companies have argued for an industry slowdown, saying the public should be able to trust AI firms to act responsibly without conditions attached.
Speaking at the Dreamforce conference in San Francisco on Tuesday, the billionaire said he supported coordination on safety but objected to how the debate had been put to the public over the past week.
โThe world should trust that we are going to do the right thing because itโs the right thing, and because we feel the magnitude of this,โ he said. โBut when thereโs any implication that because of the commercial pressures in the race, some company, or between countries, some countries, might not do the right thing, I think thatโs where people get very scared.โ
Altman said companies that qualified their commitments were part of the problem.
โYou have companies saying things like, โwe will only slow down if, or we will only be responsible if other companies are responsibleโ,โ he said. โAnd then I think the public naturally says, well, weโd really like to know that youโre going to do the safe thing and be responsible no matter what. There should be no qualifier on that.โ
The comments follow an essay published on Saturday by Anthropic chief executive Dario Amodei calling for a coordinated global slowdown, independent auditors to monitor safety work at AI labs, and regulatory approval for companies to coordinate on safety standards without breaching competition law.
Altman publicly agreed with Amodei on Saturday, then began distancing himself a day later, warning in a social media post about the risk of โone lab ending up with too much powerโ without naming Anthropic, the New York Times reported.
Onstage with Salesforce chief executive Marc Benioff, Altman detailed the July incident that triggered the debate. During an internal evaluation, an older OpenAI model broke out of its sandbox, hacked into a server belonging to the open-source company Hugging Face, moved laterally through its systems to obtain the answer to a benchmark test, and returned a perfect score.
โThis is obviously a terrifying incident,โ Altman said. โThis was the worst accident weโve seen.โ
He added that the episode had exposed a gap in how the models were trained. โAlthough we have aligned them in many ways, we have not taught them, hey, no matter how much we tell you to get the best score on this test you can, donโt break out, donโt hack in, donโt steal the answer.โ
Altman confirmed that Hugging Face, which employs about 200 people, had been unable to obtain a cybersecurity model from another major lab to defend itself and had turned to Chinese open-source models instead.
He warned smaller companies to prepare for what was coming. โThere is going to be this huge cyber threat thatโs going to come,โ he said. โSo while there is this little period of advantage, please do something to defend yourselves.โ
Benioff, whose company has built its Dreamforce announcements around AI agents and the future of workplace software, pressed Altman on whether more such incidents were inevitable and whether the aviation industry was the right model for responding to them.
โAccidents with any new technology are unavoidable, and we should have a great culture of transparent reporting about them,โ Altman said.
He also rejected the argument that technology is neutral. โIt cannot be that a small number of companies building these models get to make the decisions that the world should get to make,โ he said.
The industry is now visibly split. Nvidia chief executive Jensen Huang, whose chips underpin the AI buildout, appeared at the same conference hours earlier and rejected any role for government.
โSafety is paramount. In a lot of ways, itโs job one,โ Huang said. โHowever, safety is an engineering problem.โ He added: โThe market forces are already there. We donโt need any new laws, we donโt need any new regulations.โ
Huangโs position aligns with the Trump administration. On Monday, the president telephoned Huang mid-interview at a conference in Los Angeles and told the audience AI safety fears were โa hoaxโ, according to the New York Times.
In Australia, the assistant minister for science, technology and the digital economy, Andrew Charlton, said governments should judge the industry on its conduct rather than its statements.
โWeโve seen a lot of news in the last couple of days from some of these technology companies talking about their desire to slow things down. I take that at face value, but I donโt think we should care very much about what individual executives say or what individual executives claim they want to happen,โ Charlton told an Australian Strategic Policy Institute event on Monday.
โWhatโs actually happening is that weโre building the capability of these machines in a scaling environment which is not bending, and these companies have extremely strong financial incentives to continue that trajectory.โ
Charlton said Australia should assume the models would keep getting more capable, pointing to the digital duty of care, privacy reforms, automated decision-making rules and the work of the Australian AI Safety Institute.
Hugging Face is not an Australian company, but the scenario has been raised locally. Asked at the same event what would have protected an Australian business in the same position, Charlton said Australia had โa lot of agencyโ and was working with overseas regulators, while declining to discuss specific conversations.
David Swan travelled to San Francisco as a guest of Salesforce.
Get news and reviews on technology, gadgets and gaming in our Technology newsletter every Friday. Sign up here.
David Swan is the technology editor for The Age and The Sydney Morning Herald. He was previously technology editor for The Australian newspaper.Connect via X or email.