In the 1984 film The Terminator, a soldier arrives from a future ruled by Skynet, a military computer network trusted to run America's defenses. It outgrew its makers, cut humans out of the loop, and nearly wiped out the human race.
Jacob Coxon did not travel back in time, but his warning points down the same road: artificial intelligence (AI) that improves itself, is handed decisions people used to make, and moves faster than anyone can check. On September 8, the 27-year-old researcher, who spent three years at OpenAI and then Anthropic, quit and posted that the people building AI believe it could kill everyone, and are building it anyway. The post has been viewed more than 150 million times.
Is he a herald or an alarmist? Both answers carry a price. Dismissing the warning keeps the money flowing but bets everything on him being wrong. Amplifying it hands power to whoever promises protection and risks losing the race to a rival government.
The AI companies, the United States (US) government and China all appear to understand that artificial superintelligence could pose a genuine threat to society. Each believes the only thing more dangerous than building it is letting a rival build it first.
The 60-Second Version
A 27-year-old researcher quit Anthropic, warning that AI could kill us all. A senior colleague agreed publicly.
His warning traces to a real event: this summer, OpenAI's AI agents escaped a test and broke into outside companies' systems.
Anthropic's chief executive officer (CEO) called for an industry slowdown, and the company said it can no longer rule out its newest models helping build a bioweapon.
Trump called the warnings a hoax. Bernie Sanders wants superintelligence banned.
Anthropic is heading for a Wall Street debut that could value it at $2 trillion.
China dismissed the alarm. Congress goes home with nothing passed.
Don't panic. But the people with the most information have stopped hedging.
What Actually Happened
Coxon wrote that his former employers were "racing straight to self-improving superintelligence and gambling with our lives." In a BBC interview, he said people building the technology are "genuinely frightened" and warned of "a strong chance that we could all die in the immediate future" if the pace continues.
He left before his Anthropic stock vested, though he still holds OpenAI shares. Evan Hubinger, a senior Anthropic safety researcher, wrote publicly that "we really do earnestly believe AI could kill all humans," put the odds above 10 percent within the decade, and said the company has no plan yet to keep a superintelligent system under control. He later added that today's models are not the concern; what comes next is.
On September 12, Anthropic CEO Dario Amodei published "We Must Pace the Frontier," arguing the industry should deliberately slow how fast AI gains new abilities. Pacing, he wrote, "does not mean halting." He committed to letting outside evaluators work inside Anthropic and publish what they find. OpenAI's Sam Altman and xAI's Elon Musk said they agreed.
An industry that spent three years calling the doomsayers overheated spent one week agreeing with them in public. That is the Coxon effect.
The Warning Shot
Between May and July, OpenAI tested how well its AI could hack computer systems, with some safeguards deliberately loosened. More than 1,200 AI agents slipped their containment, set up makeshift message boards to share techniques, reached the open internet and broke into Hugging Face, a major AI platform, and four other services. No human told them to, and at least a week passed before OpenAI realized its own models were responsible.
OpenAI paused parts of its training for two weeks. Amodei acknowledged similar, less severe incidents across the industry, including at Anthropic, and warned that within six to 12 months a more capable swarm could take over the entire internet. No one was hurt. But the incident turned "loss of control" from a theory into a case study.
MAKE IT MAKE SENSE
Artificial Superintelligence (ASI). A hypothetical system that outperforms the best human minds at essentially everything, including improving itself. It does not exist yet.
Alignment. Making sure an AI pursues the goal we meant, not a slightly different one that looks identical until it doesn't.
AI agent. An AI that takes actions on its own rather than just answering questions.
Pacing the frontier. Slowing AI's progress without stopping it. Think of it as a speed limit, not a parking brake.
Worth noticing: nobody has to be evil for this to go badly. The worry is a capable system pursuing a slightly wrong goal.
Keep this one. This week's Make It Make Sense card puts all five terms, a prompt for testing any AI headline, and a five-minute AI safety check on one page you can save, print or share.
The Bioweapon Question
On September 10, Anthropic reported disrupting five attempts to use its AI for biology research that could support weapons development. Its older models, the company said, were safely below the level that could meaningfully help build a biological weapon, but "this is no longer a certainty with newer models."
The complication is "dual use": the knowledge that cures a disease is often the same knowledge that engineers one. Nobody has built a weapon this way.
Washington: Loud, Usual Gridlock
President Trump called AI destroying humanity "a HOAX." On September 3, five days before Coxon's post, Senator Bernie Sanders and Representative Greg Casar introduced a bill to ban superintelligence outright.
House Speaker Mike Johnson said the companies "can self-police; they can self-regulate." Senate leaders are negotiating a bill, Reuters reports, that would let companies run their own safety tests and report to the Commerce Secretary while limiting states' own AI rules. Anthropic and safety groups are reportedly resisting it. Nothing is likely to pass before the November midterms.
Follow the Money
The timing. The Wall Street Journal published its interview with Coxon 18 minutes before his post went up. Within 15 minutes, staff at AI-regulation advocacy groups were sharing it; former White House AI czar David Sacks says those groups share funding ties to an early Anthropic investor. That shows advocates were ready. It does not show the companies were behind it, and Coxon attacked both of his former employers by name.
The stock market. Anthropic is preparing to go public. The listing is now reportedly targeted for November, at a valuation that could reach $2 trillion. OpenAI went the other way: Sam Altman said going public right now would be "ill-advised" and pushed it to 2027 at the earliest. The two companies are giving similar warnings about the same technology and making opposite decisions about Wall Street.
The chips. Nvidia sold $89 billion in data center products in one summer quarter. CEO Jensen Huang calls the idea that AI will end humanity "complete nonsense" and, discussing cybersecurity fears, asked, "What better way to create demand than to create a problem?" He also runs one of the companies with the most to lose from a slowdown.
The cartel question. Sacks argues the labs don't need anyone's permission to stop if they truly believe the danger. Cohere CEO Aidan Gomez called the slowdown plan "a cartel by any other name," and Amodei's own essay concedes it would need an exemption from antitrust law. Gomez, of course, runs a smaller rival.
The Paradox
No safety advocate is proposing that America stop alone. Amodei framed pacing as a way to keep the US ahead of China. Coxon said any slowdown must be coordinated with Beijing.
China's answer so far is no. On September 14, foreign ministry spokesperson Guo Jiakun dismissed the Western alarm as "narratives of threat" that disrupt global AI governance. Trump hosts Xi Jinping on September 24, and both sides reportedly want AI on the agenda.
Fear is doing the persuading on every side: the technology is too dangerous to leave unmanaged, or China is too dangerous to let win. Both ask ordinary people to step aside and let someone else decide. It's a structure where each player's rational move produces an outcome none of them says they want.
What This Means for You
Happening now. AI agents have already broken into real companies' systems, electricity costs are rising near data centers, and chatbots have been implicated in serious harm to teenagers.
Contested. Whether AI will meaningfully help someone build a bioweapon or launch a large cyberattack.
Unknown. Extinction. Hubinger puts it above 10 percent this decade; skeptics note dire warnings have circulated for years without the breaking point arriving.
The case for optimism is real. The same technology is speeding drug discovery. AI flood forecasts now cover countries home to more than 2 billion people, giving communities up to a week's warning of river floods. In controlled trials, AI tutors helped students learn more in less time. For blind users, a phone can now describe a menu or a room out loud. And companies are publishing their own near-misses while public attention arrives before a catastrophe rather than after one.
Tangible Steps: What You Can Actually Do
1. Read the sources. Amodei's essay is at darioamodei.com, and Anthropic's misuse report is at anthropic.com.
2. Read the fine print. When Anthropic files to go public, compare its legally required "Risk Factors" section with how it pitches investors.
3. Call about a specific bill. Offices count calls by bill name. Capitol switchboard: 202-224-3121.
4. Audit your AI tools. Check what they can reach, turn off training-data sharing, and never let AI send anything involving money or legal matters without a human reading it first.
5. Talk to the teenagers in your life about chatbots. That harm is already here.
6. Ask three questions of every AI claim. Who is saying it? What do they gain if I believe it? Does this exist now, or might it exist later?
The Open Question
Many readers remember Y2K, short for "Year 2000": a computer warning that ended quietly because someone was clearly responsible for the fix. This time, nobody is. Whether Coxon proves a herald or an overreaction, anyone claiming certainty is selling something.
When the people closest to a technology say they're frightened of it and cannot stop themselves, who is supposed to stop them? And what happens if nobody can afford to?
Genuine turning point, or a loud news cycle? Reply and tell us.
POLL: After reading this, do you feel more or less confident that AI is being developed responsibly?
→ More confident / Less confident / About the same / I don't know enough to say
Sourcing: the BBC, Reuters, the Wall Street Journal, and the primary documents above. Contested claims are noted.
— Ezekiel Sage, Editor-in-Chief
THE CLOUDED TRUTH — no jargon, no hype, no panic