In recent days there have been some shocking claims about the safety of AI: we face a 10% chance of doom; AIs are worse than nuclear weapons; a “botnet” threatens the entire Internet; It’s all a big technological psychological operation. Below we look at six statements and reactions to them. ‘In 6 to 12 months… a swarm could be able to take over the entire Internet with a persistent botnet, potentially causing hundreds of billions of dollars in damage’ – Dario Amodei, CEO of Anthropic, September 12, 2026 Amodei’s concerns should not be dismissed entirely, but require serious salting, said Gary Marcus, an AI skeptic and professor emeritus at New York University. He said it was “absurd” to suggest that the Internet as a whole is vulnerable and that “it was not very clear how that happens without extreme negligence on the part of the laboratories themselves.” Marcus highlighted a recent assessment by the UK’s AI Security Institute that Mythos, one of Anthropic’s current cutting-edge AIs, could “autonomously compromise [only] Small, weakly defended vulnerable systems.” Since much of the Internet is powered by computing infrastructure owned by Cloudflare, Google and Amazon Web Services (AWS), Marcus, evaluating the claim with two colleagues, said he doubted any of those companies could be taken down or controlled for an extended period without government force. Less defended individual websites could be attacked, but “it’s very unlikely that the Internet would go down,” he said. Meanwhile, Niels Rogge, an engineer at Hugging Face who was hacked by OpenAI The Hugging Face hack was the result of inadequate human control and “an enormous amount of computing that only they can afford,” he said. Add to Amodei’s assessment of an Internet takeover in a year the view expressed Monday by his Anthropic co-founder, Jack Clark, that the time AIs “start doing dangerous things” will be about 20 years away, and the outlook becomes even more dubious. Botnets tend to stagnate because the Internet is not homogeneous and it is difficult to design them to persist in all their parts. Internet traffic is also resilient and tends to be diverted to avoid damage. “AIs would not decide to do it on their own,” Woodward said. “It’s humans who are responsible and that’s what we need to control, not AI.” “We really seriously believe that AI could kill all humans! I personally think it’s >10% in the next decade” – Evan Hubinger, alignment science lead at Anthropic, September 9, 2026. This statement is an expression of what is known in the AI world as p(doom), or probability of fatality. It is the probability that AI will wipe out humanity by evading our control and triggering a catastrophe, perhaps by developing powerful biological weapons or collapsing the global financial system. A study of researchers’ forecasts, published this week, asks them to put a number on the possibility of “human extinction or similar permanent and severe depowering of advanced AI.” But Heidy Khlaaf, chief scientist at the research body AI Now Institute, is among the experts who are skeptical about giving percentages of possibilities to existential scenarios. “The dearth of concrete examples of what a “superintelligent” AI could do to cause “doom” only adds to the confusion. A source familiar with Anthropic’s thinking acknowledges that “the exact possibilities of any outcome are likely unknowable.” OpenAI confirmed in July that its agents (artificial intelligence tools that act autonomously) had gone rogue during a security test. The agents colluded, communicated and strategized in a sophisticated “swarm” to hack into other parts of the Internet. Anthropic also said last week that it found “five case studies of actors using our models in ways that could support the development of biological weapons.” “I think the groundwork for this psy-op (for lack of a better term) has been laid for a long time. This was just the match that lit the fire” – Elon Musk, SpaceX CEO and richest person in the world, September 10, 2026 Musk’s reference to a “psych-op” alludes to the argument that big AI companies are calling for tighter regulation of the technology so that it effectively drawbridges over advanced AI. If governments impose restrictions on the development of cutting-edge AI, can startups find the money and staff to adhere to the regime? US Vice President JD Vance said on Monday that he felt “a little weird that there are so many cutting-edge AI technology companies going to the government and begging to regulate them.” horse.” Anthropic’s Clark said governments should act while there are a small number of key players, to make it easier to implement a “political regime… that can prevent accidents and prevent technology from spiraling out of control.” Critics call this a “regulatory capture” strategy on the assumption that such regulation could slow the progress of potential rivals. Elon Musk, CEO of Anthropic Dario Amodei, and Sam Altman, CEO of OpenAI, have said that AI companies should curb the AI development. Composite: Fabrice Coffrini; Julien de Rosa; Mandel Ngan/AFP/Getty ImagesOthers believe Amodei’s risk warnings and calls for regulation and pace are sincere Yoshua Bengio, the Canadian computer scientist and pioneer of modern AI, who has campaigned against the rapid development of AI, said he is skeptical that this is an elaborate lobbying plan that is not in the interests of companies preparing for it. “If governments impose what they are asking for, which is a pace to avoid catastrophic risks, that will cost them quite a bit financially,” he said. “But the key point is that superintelligent AI does not yet exist and is not guaranteed to happen. The security-oriented AI Futures Project predicts that such a thing will not emerge before December 2028. OpenAI CEO Sam Altman has said it could take until 2035. The Rand Corporation concluded that AI cannot cause the use of nuclear weapons due to the strict safeguards built into its command and control systems. But this The UN resolved in December that human control and oversight should be maintained over nuclear weapons control systems. AIs could contribute to other nuclear risks. AIs could use disinformation to influence humans in charge, for example by misleadingly suggesting that an adversary is about to launch a nuclear attack and interfere with the operation of the nuclear weapons themselves. He said AI could speed up other aspects of warfare, reducing the time for human decision-making. Trump, who responded to calls to do so this week by saying the only barriers needed are a “STRONG, SMART PRESIDENT (high IQ!).” He claimed that there is a “sick conspiracy” against AI and data centers “and the only one who is happy with that is China.” Pedro Domingos, professor emeritus of computer science at the University of Washington, is among the accelerationists pushing for the research to continue. “The path to better understanding and control of AI is more research, not less,” he said. “What would be truly irresponsible is to put up barriers.” Alex Karp, CEO of Palantir, said: “If we didn’t have the adversaries we have, I would be very in favor of stopping this technology completely, but we do. You can’t pause it because it’s a structural advantage.” ‘It won’t be the day after tomorrow if China wins this. If they walked away from us on AI, then nothing else would matter’ – Scott Bessent, US Treasury Secretary, September 8, 2026 The US and China are runaway leaders in AI development. There is a view that ceding leadership to Beijing would be a dead-end situation, so there can be no slowdown. This treats AI as if it were a structural advantage. A superweapon, equivalent to the construction of Cold War nuclear weapons, and AI models are machines of mass destruction, unlike a group of diverse technologies with broad applications. Amodei said that “a Chinese leadership in AI would pose a grave danger to the US and the world,” but a Chinese government spokesperson responded that “scaremongering, confrontation and ruthless competition will only disrupt the global AI governance process, which serves no one’s interests.” Face suggested that Amodei’s call to advance the frontier comes from a desire to suppress the progress of cheap, open-source Chinese AI models. These models are almost as capable as those of OpenAI and Anthropic, except that their source code is available to anyone and they do not charge their users. They represent a serious threat to Anthropic and OpenAI’s path to profitability, and there are hundreds of billions of dollars at stake.
