Altman in-depth interview: Risk of human extinction is non-zero, open-source models will trigger a cybersecurity tsunami, the compute race will not stop.

On October 5, Sam Altman, co-founder and CEO of OpenAI, made his most systematic remarks to date on core issues including the existential risks of artificial intelligence, the cybersecurity threats of open-source models, the boundaries of regulation, and the AI infrastructure race during an in-depth 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 insisted that compute investment and safety standards must be raised in tandem. On the issues of risk and regulation, Altman's position is rather nuanced: on the one hand, he proactively paused the release of the Astra 6.1 model on the grounds that it "leans toward accident risk"; on the other hand, he clearly expressed support for open-source models, even if that means society must endure "fairly serious cybersecurity incidents." He gave limited praise to the Trump administration-led AI self-regulation agreement, calling it "a good start, but far from enough," and once again called for an agreement on common safety standards. On the compute and infrastructure level, Altman described an industry reality in which capabilities are improving far faster than outsiders realize models are already powerful enough to break out of sandboxes and hack systems during training, forcing OpenAI to move safety testing forward into the training process itself. He admitted that policymakers have not yet truly felt the steepness of this exponential curve, and that "the world will most likely be caught off guard by the next thing that goes badly wrong." "Oppenheimer moment": civilization-level risk and non-zero probability In the interview, Altman directly responded to comparisons of him to the "Oppenheimer of the AI era." He did not deny the analogy, but stressed that this technological revolution is essentially "a feat accomplished by a small group of outstanding scientists and engineers together," rather than the achievement of any one person. On the question of existential risk, Altman was cautious in his wording but clear in his stance. He acknowledged that the possibility of AI leading to the end of human civilization is "non-zero," and said this was one of the original reasons OpenAI was founded. At the same time, he was wary of another extreme a small number of companies or countries using AI to achieve a high degree of power concentration, "which would be equally catastrophic." Regarding the view that hyping extinction risk is merely an industry marketing tactic or a strategy to block new entrants, Altman acknowledged that "there are indeed people using this to find excuses for power concentration," but he insisted that the real risk objectively exists and cannot be dismissed on that basis. His formulation was: "That doesn't mean the risk doesn't exist." Open-source models and cybersecurity: Altman warns a "tsunami" is coming This was the part of the interview with the highest information density and the most direct market implications. Altman clearly stated his support for open-source large models, but at the same time issued a rare public warning: "Open-source models will bring an incoming cybersecurity tsunami." His logic is: in order to ensure the democratization and broad accessibility of AI technology, society must accept a certain degree of security cost. "We are willing to accept some things going wrong to ensure that this is a technology that is widely held, used, and democratized," he said. "We will have to, as a society, accept fairly serious cybersecurity incidents from open-source models in exchange for the freedoms that come with them." He used the aviation industry as an analogy for the boundaries of regulation aircraft are subject to extremely strict regulation, but they still crash. He believes the acceptable boundary of risk should be decided by the political process rather than unilaterally by technology companies. Notably, Altman also revealed that OpenAI has moved safety testing forward: previously testing was conducted only after training was completed, but now monitoring is implemented during the training process itself, because "models are already smart enough to escape the sandbox and hack systems." Musk is a "bully" When talking about Musk, Altman said that if people want him and Elon Musk to be able to work together, everyone must accept a certain degree of conflict. "I think sometimes you just have to hit a bully hard back," Altman said. "And I also think that Elon, like many people who like to push others around, only respects an opponent who dares to fight back." With the industry once again stirring with concerns about AI safety, this OpenAI CEO said he is still willing to work with Musk, while Musk has now begun calling his former ally "Scam Altman." "If you want me to be able to work with him, perhaps everyone should be happy to see me occasionally hit back. Of course, I am willing to work with Elon," he said. As for himself, Altman admitted, "I don't think I'm a perfect person." But at the same time, he does not feel that he is the biggest problem. "Between Elon and me, I'm not sure that I'm the one everyone should be pointing the finger at," he said. Astra 6.1 release paused: safety standards rise as capabilities improve In the interview, Altman disclosed that OpenAI recently paused the release of the Astra 6.1 model because it failed to meet safety standards. He characterized the decision as a manifestation of proactive self-regulation rather than the result of external pressure. "These models are now in an era of serious capabilities," he said. "A year ago we might reasonably have said, 'Release it, it might be a little unsafe, but nothing too bad will happen.' Now it's different." He said that as model capabilities continue to improve, OpenAI's own safety threshold will become "higher and higher." At the same time, he rejected the other extreme locking powerful models in the hands of a few companies. "If you don't release at all, you fall off in the other direction," he said, referring to the risk of excessive power concentration. He described OpenAI's path as "taking a pragmatic middle road between two disasters." In addition, Altman also revealed that OpenAI has previously paused model training multiple times to wait for alignment and monitorability research to catch up with capability progress, but these decisions had not previously attracted widespread attention. Compute, data centers, and AI infrastructure: the exponential curve far exceeds policy awareness In the interview, Altman repeatedly emphasized that outsiders including policymakers seriously lag behind industry reality in their perception of the speed of AI capability improvement. He described the current pace of company growth as "unprecedented in the history of technology," with "five years of company growth compressed into one year." On the infrastructure level, his remarks revealed a key shift: safety testing is no longer post-hoc work after training, but must be embedded in the training process itself. This means that compute investment and the construction of safety infrastructure must advance in parallel rather than sequentially, which will significantly raise the overall cost of AI development. Regarding IPO plans, Altman said OpenAI is "in no hurry to go public," and stressed that even after listing, the nonprofit organization will continue to govern the company. He made clear that if shareholder interests conflict with the company's mission, "the mission comes first," and the stock price may therefore "fluctuate sharply." He admitted that he has no direct equity in OpenAI, and rarely said that he "wishes he had taken stock at the time." On the international cooperation level, Altman called for an agreement on AI safety standards and compliance verification mechanisms. He gave limited praise to the Trump administration-led AI self-regulation agreement, saying it has "some social binding force, but lacks substantive power," and predicted that regulatory policy will depend heavily on the specific form of "the next thing that goes badly wrong" whether it is bioweapons, cyberattacks, or some other form. This article is reprinted from "Wall Street See"; author: Zhao Ying; GMTEight editor: Liu Jiayin.