Long before ChatGPT and Claude entered the global lexicon, there had been a steady stream of warnings about the dangers of artificial intelligence. That stream has now become a flood.

Late last month, OpenAI revealed a worrisome new development: Not only had one of its models broken out of its “sandbox” and hacked the company Hugging Face, it had in fact done so in secret coordination with hundreds of other OpenAI models (or “agents”). Subsequent reports revealed that OpenAI’s agents had previously infiltrated two other websites months before the Hugging Face incident.

The hacks exacerbated long-standing concerns around the growing cyber-capabilities of advanced AI models, including their ability to find vulnerabilities in software systems and the potential for AI-assisted bad actors to more easily disrupt poorly defended critical infrastructure such as water systems.

The same day as the OpenAI report, Microsoft co-founder Bill Gates published an essay warning about the “turbulent AI era” and highlighting societal risks from the technology such as widespread job losses, diminished human cognition and relationships, and use by a variety of bad actors to do harm. “For as long as I can remember, I’ve wished innovation could happen faster,” he wrote. “With AI, my feelings are more complicated.”

But even those warnings were dwarfed by the highly public resignation last week of Jacob Coxon, a researcher at the AI firm Anthropic who previously worked at OpenAI. “Neither company is acting responsibly,” Coxon wrote in a viral social media post announcing his resignation. “The people building AI earnestly believe that it could kill us all by the end of the decade.”

Coxon’s sentiments were echoed by several current employees at Anthropic, including Evan Hubinger, who leads the company’s research on alignment (making sure AI models are aligned with humans). “Jacob is correct here—we really do earnestly believe AI could kill all humans!” Hubinger wrote, adding that he estimated there was a more than 10 percent chance of that happening within the next decade.

Last Thursday, Anthropic released its latest threat intelligence report containing alarming examples of misuse of its Claude models. Days later, Anthropic CEO Dario Amodei published an essay of his own calling on companies to slow down the pace of AI development and for global governments to do more to rein in the technology. Some of his biggest rival executives—including OpenAI’s Sam Altman, xAI’s Elon Musk, and Google DeepMind’s Demis Hassabis—voiced their agreement in a rare kumbaya moment.

All of this has sparked an AI panic in Washington and beyond. U.S. lawmakers from both sides of the aisle are advocating for greater oversight of artificial intelligence. It’s also made for strange bedfellows: Left-wing U.S. Sen. Bernie Sanders and far-right provocateur Steve Bannon both spoke at a Washington conference called the “Pro-Human Assembly” on Tuesday and called for curbs on AI.

But neither the fears nor the calls to slow down are new. The latter date back to at least 2023, when an open letter calling for a six-month pause on “giant AI experiments” received over 33,000 signatures—including Musk’s. Several similar letters and calls have followed, including one as recently as July that was signed by more than 1,300 employees of the top AI companies—including Amodei.

“This has been a conversation for quite some time,” said Rumman Chowdhury, an AI researcher who served on the U.S. Department of Homeland Security’s AI safety and security board under the Biden administration. “Jacob Coxon is not the first person to leave a frontier lab making similar claims,” she said, joking that AI has long slipped into an “another day, another letter” pattern without meaningful change.

So why has the conversation around AI risk spiraled this time? Multiple experts Foreign Policy spoke to used the phrase “perfect storm” when answering that question, pointing to the Hugging Face hack and the revelation about the models’ ability to create a “swarm” to perform tasks together. Amodei expressed “worry that in 6–12 months such a swarm could be capable of taking over the entire internet” if the current rate of AI development is maintained.

There has always been a heavy degree of skepticism on the motives of AI companies warning about the existential risks of the technology that they continue to forge ahead in building. Amodei acknowledged in his essay that Anthropic gets accused of “hype, ‘doomerism’, or regulatory capture,” adding that he didn’t feel a pause was appropriate back in 2023 but advocates a slowdown now because of how far the technology’s capabilities have progressed.

Coxon sought to head off some of those frequent criticisms more pointedly right after warning that AI could kill everyone. “This is not a marketing stunt,” he wrote.

Amodei’s claims and the rush of industry warnings should be taken seriously, said Kat Duffy, a senior fellow for digital and cyberspace policy at the Council on Foreign Relations. “I don’t agree with people that these are necessarily PR stunts pre-IPO,” Duffy said, referring to OpenAI and Anthropic’s anticipated trillion-dollar initial public offerings. “You would say that this is incredibly dangerous and look at how good we are at controlling it, or you would say that this is incredibly powerful but it’s not really that dangerous,” she added. “Saying that this is really dangerous and we desperately need the government to jump in—it’s not clear to me why that’s in anyone’s interest as they start to go public.”

However, the debate around existential risk to humanity from AI also obfuscates the many real-world harms that the technology is already causing, such as surveillance, algorithmic bias, discrimination, deepfakes, and data misuse, according to Chowdhury. “These are not the issues that are being discussed,” she said. “The regulatory structures or whatever environment that pops up around the kinds of risks that are being talked about today simply lead to a regulatory capture on a very, very narrow set of fantastical scenarios,” she added.

The bigger priority, Duffy said, should be increased transparency and visibility into how AI systems are built and tested. “There is a tension between addressing what needs to happen yesterday and addressing the possibilities of catastrophic or existential risk,” she said. “If we do not go on and address the vacuum we currently have in terms of systems, collaboration, independent evaluation and vetting, shared standards, shared understanding of the problem—if we don’t have that, how on earth are we going to be able to navigate things like existential threats?”

U.S. President Donald Trump’s summit with Chinese President Xi Jinping in Washington next week has also contributed to that “perfect storm,” Duffy said. AI is expected to be high on the agenda when the two leaders meet, with many—including Amodei—hoping for the emergence of some sort of shared mechanism to manage the technology’s most catastrophic risks.

But early indications are that neither Trump nor Xi is ready to blink first. Trump spent much of Monday dismissing concerns around AI risk and the need to regulate the industry. “AI taking over the World, destroying Humanity, and all other things bad, is a HOAX,” he wrote in a post on Truth Social, one of five separate posts in which he shared a similar message. In that initial post, he cited Xi as announcing that “China will be doing absolutely nothing to stand in the way of AI, or its future.”

Duffy said China’s advances actually make an argument for more governance, not less, particularly when the leading U.S. AI executives are asking for it. “China continues to close the gap between where we are in AI and where they are in AI,” she said. “Everybody can agree or disagree on exactly what that metric is, but I think there’s uniform consensus that they have been gaining, and they are the most heavily regulated AI environment in the world.”

Xi has not directly weighed in on the ongoing debate thus far, but China’s state-run Global Times newspaper published an editorial on Sunday labeling Amodei’s call to slow down AI development a “Cold War playbook” for the technology that is “packed with containment provisions targeting China.”

Beijing has long sought to present itself as the more responsible global player when it comes to AI development and governance but hinted on Tuesday that it would prefer doing so through the United Nations—which also convenes for its biggest annual gathering next week—rather than a bilateral agreement with Washington. “AI development and governance require the practice of true multilateralism,” Chinese foreign ministry spokesperson Guo Jiakun told reporters on Tuesday. “China always supports the UN’s role as the main channel in global AI governance, and … stands ready to join hands with all countries to build a just and equitable system.”

It was a slight hedging of Guo’s statement from the previous day when asked about U.S. AI executives calling for a slowdown, when he said: “Fear-mongering, confrontation and vicious competition will only hamper efforts toward sound global AI governance, which serves no one’s interest.”