Altman's in-depth interview with Vanity Fair: The risk of human extinction is non-zero, open-source models will trigger a cybersecurity tsunami, and the computing power competition will not stop.
On October 5, Sam Altman, co-founder and CEO of OpenAI, made his most systematic statement to date on core issues including existential risks of artificial intelligence, cybersecurity threats posed by open-source models, regulatory boundaries, and the AI infrastructure race during an in-depth exclusive interview with *Vanity Fair*.
He acknowledged that the probability of AI leading to the end of human civilization is "non-zero", while warning that open-source large models will trigger a "cybersecurity tsunami", and insisting that computing power investment and safety standards must be improved in sync.
On the issues of risk and regulation, Altman's stance is rather nuanced: on one hand, he voluntarily suspended the release of the Astra 6.1 model on the grounds that the model is "prone to accident risks"; on the other hand, he explicitly expressed support for open-source models, even if it means society has to bear "quite serious cybersecurity incidents".
He gave a limited evaluation of the AI self-regulation agreement led by the Trump administration, calling it "a good start, but far from enough", and once again called for an agreement on common safety standards.
At the level of computing power and infrastructure, Altman described an industry reality where the speed of capability leap far exceeds external perception — models are now powerful enough to break through sandboxes and invade systems during the training process, which forces OpenAI to move safety tests forward to the training phase itself.
He frankly admitted that policymakers have not yet truly felt how steep this exponential curve is, "the world will most likely be caught off guard by the next major mishap."
01
"Oppenheimer Moment": Civilization-level Risks and Non-zero Probability
In the interview, Altman directly responded to the public's comparison of him to "Oppenheimer in the AI era". He did not deny this analogy, but emphasized that this technological revolution is essentially "a feat accomplished together by a small group of outstanding scientists and engineers", rather than the achievement of a single individual.
On the issue of existential risks, Altman used cautious wording but held a clear attitude. He acknowledged that the possibility of AI leading to the end of human civilization is "non-zero", and said that this is one of the original intentions for OpenAI to be founded. He also warned against another extreme — a small number of companies or countries achieving highly centralized power with the help of AI, "which is equally catastrophic".
Regarding the view that exaggerating the risk of extinction is nothing more than a marketing tactic of the industry or a strategy to block new entrants, Altman admitted that "some people are indeed using this as an excuse to seek power centralization", but he insisted that the real risks exist objectively and cannot be denied for that reason. His exact words were: "That doesn't mean the risk doesn't exist."
02
Open-source Models and Cybersecurity: Altman Warns a "Tsunami" is Approaching
This is the part of the interview with the highest information density and most direct market implications. Altman explicitly expressed support for open-source large models, but at the same time issued a rare public warning: "Open-source models will bring an upcoming cybersecurity tsunami."
His logic is: To ensure the democratization and broad accessibility of AI technology, society must accept a certain degree of security cost. "We are willing to accept that something goes wrong to ensure that this is a widely held, used and democratized technology," he said, "We as a society will have to accept fairly serious cybersecurity incidents from open-source models in exchange for the freedom that comes with it."
He used the aviation industry as an analogy for regulatory boundaries — aircraft are subject to extremely strict regulation, but they still crash. He believes that the acceptable risk boundary should be determined by the political process rather than unilaterally by tech companies. It is worth noting that Altman also revealed that OpenAI has moved safety tests forward: previously tests were only carried out after training was completed, now monitoring is implemented during the training process, because "models are already smart enough to escape the sandbox and invade systems".
03
Musk is a "Bully"
When talking about Elon Musk, Altman said that if people hope that he and Elon Musk can cooperate, everyone must accept a certain degree of disputes.
"I think there are times when you have to hit back hard against a bully," Altman said, "And I also think that Elon, like many people who like to bully others, will only respect opponents who dare to fight back."
As the industry is once again raising concerns about AI safety, the OpenAI CEO said he is still willing to cooperate with Musk, who now refers to his former ally as "Altman the liar".
"If people hope that I can work with him, maybe everyone should be happy to see me fight back occasionally. Of course, I am willing to cooperate with Elon," he said. As for himself, Altman frankly admitted "I don't think I'm a perfect person". But at the same time, he doesn't think he is the biggest problem.
"Between Elon and me, I'm not sure I'm the one everyone should be targeting," he said.
04
Astra 6.1 Release Suspended: Safety Standards Improve as Capabilities Rise
Altman disclosed in the interview that OpenAI recently suspended the release of the Astra 6.1 model because the model failed to meet the safety standards. He characterized this decision as a manifestation of voluntary self-discipline rather than the result of external pressure.
"These models are now in the era of serious capabilities," he said, "A year ago, we could reasonably say, 'Release it, it might be a little unsafe, but nothing too bad will happen.' It's different now." He said that as model capabilities continue to improve, OpenAI's own safety requirement threshold will become "higher and higher".
He also rejected the other extreme — locking powerful models in the hands of a small number of companies. "If you don't release them at all, you will fall from the other direction," he said, referring to the risk of excessive power centralization. He described OpenAI's path as "taking a pragmatic middle road between two disasters".
In addition, Altman also revealed that OpenAI has voluntarily suspended model training many times before to wait for alignment and monitorability research to catch up with capability progress, but these decisions have not attracted widespread attention before.
05
Computing Power, Data Centers and AI Infrastructure: Exponential Curve Far Outpaces Policy Perception
Altman repeatedly emphasized in the interview that the public — including policymakers — has a perception of the speed of AI capability improvement that seriously lags behind the industry reality. He described the current company growth rate as "unprecedented in the history of technology", "the company growth that used to take five years is compressed into one year".
At the infrastructure level, his statement reveals a key shift: Safety testing is no longer an after-work after training, but must be embedded in the training process itself. This means that computing power investment and safety infrastructure construction must be advanced in sync, rather than carried out sequentially, which will significantly push up the overall cost of AI development.
Regarding the IPO plan, Altman said that OpenAI is "in no hurry to go public", and emphasized that even after going public, the non-profit organization will continue to govern the company. He made it clear that if the interests of shareholders conflict with the company's mission, "mission comes first", and the stock price may therefore "fluctuate sharply". He frankly admitted that he has no direct equity in OpenAI, and rarely said "I wish I had taken the stock back then".
At the level of international cooperation, Altman called for an agreement on AI safety standards and compliance verification mechanisms. He gave a limited evaluation of the AI self-regulation agreement led by the Trump administration, believing that it "has certain social binding force but lacks substantive power", and predicted that regulatory policies will highly depend on the specific form of "the next major mishap" — whether it is biological weapons, cyberattacks, or other forms.
This article is from the WeChat official account "Hard AI", the author is Hard AI focusing on technology R&D and production, and it is released with authorization from 36Kr.