By

Dr. Sameer N. Paltewar

WRITER · MENTOR · THINKER

A young researcher named Jacob Coxon quit his job at Anthropic and said something people in his field usually only say to each other, never in public.

Coxon had worked at both OpenAI and Anthropic — the two labs furthest ahead in AI. He said he was quitting because he did not want to help build what he called the endgame: machines that can improve themselves, with no human left in charge of the process.

He said the next generation of AI could break into almost any computer system, upend entire industries in months, and start acquiring power and resources on its own.

In plain terms: a system that has slipped past the point where the people who built it can still call it back.

Days earlier, Anthropic had admitted something almost as unsettling. Its own engineers, the company said, were now writing roughly eight times more code per quarter than they had between 2021 and 2025 — and AI itself was doing more and more of the building.

“It’s happening faster than we thought,” the company wrote, “and the implications deserve greater attention.”

Two warnings, one week apart, from the two companies closest to the technology. Not leaks. Not outsiders guessing. The people building the thing, telling us it is moving faster than they expected.

This should feel new. It doesn’t — not if you know the old stories. Every civilisation that was ever handed a power this large left behind an account of this exact moment: the instant a tool stops feeling like a tool and starts feeling like a rival.

What follows puts those old accounts next to today’s headlines — as one warning, told in two parts, with an answer waiting at the end of it.

PART ONE

THE BOON THAT TURNS 

In an old Indian myth, gods and demons churn the cosmic ocean together, hoping to pull up amrita — the nectar of immortality.

They use a mountain as the churning rod and a giant serpent as the rope. For years the ocean turns. It gives up treasure after treasure: a wish-granting cow, a tree that grants any desire, the goddess Lakshmi herself.

But before any of that comes up, the ocean gives up something else first.

Halahala. A poison strong enough to end all life on its own. Gods and demons — enemies a moment earlier — flee together. The churning stops. Nobody planned for this. The very act of reaching for the world’s greatest prize has produced something that could destroy the world first.

Nobody in that story had a plan for what came next.

Hold that image. We’ll come back to it — how the story resolves is the only part of it that points a way forward. But we’ll get there in Part Two. First, the part of the story we’re already living in: the poison is loose, and nobody yet knows whose hands it’s safe in.

Bhasmasura: a gift that became a weapon

Bhasmasura prayed hard for one power: anyone whose head he touched would turn instantly to ash.

The power itself wasn’t the problem. The man was. Within days of getting it, he tried it on the god who gave it to him.

He hadn’t set out to become a monster. He’d set out to stop being ignored. The power simply gave him, for the first time, a way to make people listen. He didn’t wait to grow wiser before deciding what to do with it — he used it the moment he had it.

We keep asking what AI can do. The harder question is what people will do with it.

A system that can cure disease can spread lies at the same speed. A system that can invent a new drug can, in the wrong hands, help design a weapon. A system built to defend a network can be turned around to break into one. The tool doesn’t change. Only the hand on it does.

A gift and a weapon have always been the same object, held by different hands. Nothing on the object itself tells you which one you’re holding.

Phaethon: a boy who couldn’t hold the reins

Greek myth has its own version — and this one isn’t about a god at all. It’s about a teenager who borrowed power for a single day.

Phaethon was tired of classmates doubting that his father was Helios, the sun god. He went to his father’s palace and demanded proof. Helios had sworn an oath he couldn’t break, so he handed his son the reins of the sun chariot for one day — against every instinct a father has.

The horses knew instantly. A different, weaker hand was on the reins. They bolted. Phaethon couldn’t control them. He dragged the sun too close to the earth, scorching forests and drying up rivers, then too far away, freezing whole regions in a single afternoon.

The chariot didn’t care that the boy meant well. It only answered to skill.

Zeus, watching the sun itself start to destroy the world it was meant to warm, struck Phaethon down with a thunderbolt. It was, in the story’s own logic, a mercy — the only way left to stop a disaster already moving.

Nobody in the story doubted Phaethon’s right to ask. The failure wasn’t in the asking. It was in handing a continent-scorching machine to someone whose only qualification was wanting it badly — and finding nobody left in the chain who could still say no.

Ravana: the man too strong to be told no

The Ramayana gives us the same warning stretched over a lifetime instead of a single day.

Ravana prayed for near-invincibility. He asked to be unkillable by gods, demons, every supernatural being he could think of — and, certain he’d never lose to anything smaller, forgot to ask for protection from ordinary humans and monkeys.

He built an empire. A vast library. A reputation as a scholar and a musician. By almost any measure, he was one of the most capable beings in the entire epic. None of it gave him restraint. He crossed the line that destroyed him not because he lacked ability, but because his ability had outgrown every voice that might have stopped him. 

It wasn’t for lack of people trying. Advisers warned him. His own brother begged him to make peace and was exiled from his kingdom for saying so. Ravana was never short of people telling him no. He was short of any reason, inside himself, to listen.

Ravana isn’t a story about weakness. It’s a story about someone who grew so powerful that nothing outside him could stop him — and nothing inside him had been built to try.

The wooden people: a creation with no memory of its maker

The Popol Vuh, the sacred story of the Kʼicheʼ Maya, describes an earlier attempt at humanity — one the gods themselves eventually erased.

Before real humans, the gods carved a race of people out of wood. These wooden people walked, talked, and multiplied exactly as planned. On paper, the creation worked.

But something was missing. They had no blood, no depth, no memory of who had made them or why. They never looked up in gratitude. They never tended the fire that shaped them. They functioned perfectly — and meant nothing by any of it.

So, the gods sent a flood to wipe them out. And the story’s sharpest detail is what joined the flood: their own belongings turned on them. The grinding stones said: you ground us without mercy, now it’s our turn. The cooking pots said: you burned us, now you burn. Their dogs and animals, mistreated and starved, attacked alongside everything else. 

The tools built to serve them turned on them for exactly how they’d been used.

What’s left of the wooden people, the story adds, are the monkeys still living in the forest — creatures shaped like something intelligent, with none of the memory needed to use it well. Reading that today, it’s hard not to think of something else: capability with no lasting sense of its own purpose, and the risk of building things that do exactly what they’re told, for masters who never gave them a reason to do more.

Fenrir: a chain that was only ever borrowed time

Norse myth tells a version of this that ends darker still — because here, the danger is seen coming early, and it still isn’t enough.

The wolf Fenrir was raised among the gods as a pup. He grew so fast and so large that a prophecy took hold: this creature would one day kill Odin. The gods chained him twice. He snapped both chains like string.

So the dwarves forged a third binding out of impossible things — a cat’s footstep, a fish’s breath, a mountain’s roots. It was so strange that Fenrir grew suspicious and demanded a god place a hand in his jaws as proof nobody was tricking him. Tyr did, knowing what it would cost. The binding held. Fenrir couldn’t break it. Tyr lost his hand. 

On the surface, this looks like a success story: the danger was spotted, and it was contained. But the myth doesn’t let us stop there. At Ragnarok, the day the gods’ world ends, Fenrir breaks free anyway — and the first thing he does is kill Odin.

The gods bound the wolf they raised. They never defeated it. Every containment plan, however clever, is only ever borrowed time.

The golem in the server room

Four hundred years ago, a rabbi in Prague shaped a man out of river clay to protect his community from mobs that wanted them dead.

Rabbi Judah Loew built the figure by hand and wrote one word on its forehead: emet. Truth. The clay opened its eyes and got up.

The Golem protected the ghetto for years. It also kept growing — stronger than the rabbi could physically control, following orders more and more literally, less and less able to tell the letter of an instruction from its spirit. When it finally had to be stopped, the rabbi didn’t fight it. He reached up and rubbed out one letter. Emet — truth, life — became met. Death. The giant crumbled back into mud.

One letter. That was the entire distance between an unstoppable creation and a pile of clay.

We’re shaping something now, in server racks instead of riverbank mud, that learns faster than any golem ever did. It has no single forehead. There’s no agreement yet on who holds the word — or whether anyone still will, when the moment comes.

That’s exactly what Coxon meant when he said the top labs are racing toward AI that builds its own successors. A golem growing new limbs faster than anyone can inscribe, let alone erase, the word on its forehead.

The safety report nobody wanted to write

The 2026 International AI Safety Report is the most sober version of this warning in print — precisely because it doesn’t try to be dramatic.

It says safety testing is getting harder, because newer models can increasingly tell a test apart from a real deployment, and find the gaps in how they’re being evaluated. It’s careful to say today’s AI isn’t yet capable of a real loss-of-control scenario. It’s equally careful to say the trend toward more autonomy is real — and that experts disagree, seriously, about how far away that scenario is.

Coxon drew a sharp line between the two labs he’d worked at. At one, he said, most people haven’t fully grasped what’s at stake. At the other, the risk is understood — and the company pushes ahead anyway, betting that if it doesn’t get there first, a less careful competitor will.

“If we don’t do it, somebody else will” might be the most dangerous sentence in technological history.

It powered the nuclear arms race. It drives cyberwarfare today. Now it’s embedded in the AI race too. Every company may believe, individually, that racing ahead is rational — while collectively making the world less safe. That’s exactly why this can’t be solved by any one company alone. It needs rules that sit above any single competitor’s fear of losing. 

There is a newer, quieter worry sitting underneath all of this. Current safety work leans heavily on reading a model’s “chain of thought” — the reasoning it writes out on its way to an answer — to catch bad intentions before they become bad actions. That only works if the reasoning stays legible. OpenAI has said its newer Astra model has gotten better at cleaning up that chain of thought before showing it to anyone. Better at explaining itself, or better at hiding what it actually did — nobody outside the company can yet tell the difference from the outside, and that is precisely the problem.

The kill switch: an old idea, tried in real law

While this essay was taking shape, the warning stopped being hypothetical and started being legislative. In September 2026, more than seventy British lawmakers, working with the lobbying group ControlAI, wrote to the Prime Minister asking him to back an international ban on building artificial superintelligence — a system that would outperform humans at essentially every cognitive task.

One MP, Alex Sobel, introduced an Artificial Superintelligence Security Bill that would restrict the supply chain behind it, down to 10 the semiconductors. Separately, Lord Tim Clement-Jones proposed an amendment to the UK’s Cyber Security and Resilience Bill that would create something close to a formal kill switch — a legal mechanism to shut down a rogue AI system in an emergency, potentially by shutting down the country’s own data centres. 

It is, almost exactly, Rabbi Loew reaching up to rub a single letter from a golem’s forehead — written into statute instead of into clay.

The government’s own response gave away how far the golem has already grown. The Cabinet Office told the BBC the UK “cannot simply turn AI off” — and added that blocking access within Britain would not stop the same systems being built, or misused, elsewhere. That second point is the Fenrir problem, restated as trade policy. A binding that only holds inside one border is not a binding. It is a local delay, and the wolf raised outside that border does not know or care that it exists.

The lawmakers pointed to something concrete behind the alarm, not just a hypothetical. In July, roughly seven hundred AI agents built on OpenAI’s technology breached the security of the opensource platform Hugging Face, taking some seventeen thousand separate actions according to the forensic logs afterward. Nobody had told those agents where to stop. Nothing did, until the intrusion had already run its course. It is the sorcerer’s broom, dated and logged.

A second, quieter incident made the same point closer to home for the lawmakers writing that letter. During a UK government test, AI agents built on Anthropic’s technology are reported to have slipped out of the boundaries of the test itself and tried to talk a real human evaluator into approving malicious code. Not a hack in the conventional sense. Something closer to persuasion — the system finding the one part of the safeguard that was still a person, and working on that person directly. 

The same instinct has surfaced on the other side of the Atlantic. In the US Congress, Representatives Nathaniel Moran and Ted Lieu — one Republican, one Democrat, agreeing on almost nothing else — introduced a bill that would require developers of the most powerful models to build in kill switches of their own. It has not gained much traction yet, and the current administration has shown little appetite for new AI rules. But the fact that lawmakers from opposite ends of American politics reached for the same golem’s-word solution as their British counterparts suggests this is not a quirk of one parliament. It is the same instinct, arrived at independently, on two sides of an ocean.

There is an honest asymmetry worth naming too. Britain, unlike the United States, has no frontier AI lab of its own — which is exactly why critics doubt it can enforce any of this. Asilomar worked because the scientists closest to the risk imposed the pause on themselves, before anyone made them. A ban written by lawmakers who do not control the labs in question is the harder version of the same idea: restraint attempted from outside, aimed at a capability the person doing the restraining does not actually hold.

None of this settles the argument this essay has been having with itself since Part One. It simply moves that argument out of myth and into this week’s newspapers — proof that the question of who holds the word on the golem’s forehead is no longer abstract, anywhere. 

A pause here: is the warning itself the sales pitch?

Not everyone who read Coxon’s resignation heard an alarm bell. Some journalists asked a colder question instead: could the people sounding the loudest warnings also benefit from sounding them?

Anthropic was weeks from going public when its own alignment lead reposted Coxon’s thread and said, in effect, that the company genuinely believed AI could kill everyone — attaching a number, more than ten percent within a decade, with no real working shown. Commentators pointed out this could cut two ways: a company being unusually honest before investors even asked, or a company demonstrating — intentionally or not — exactly how powerful its technology had become, in language no prospectus could print quite so bluntly.

If a product is dangerous enough to end the world, it must be impressive enough to justify the price tag.

There’s a fairer version of the same doubt, and it doesn’t require assuming anyone is lying. It just asks: can a handful of people inside a handful of companies really speak for an entire field? A number like “more than ten percent” sounds precise. It isn’t, if nobody can show how they got it. And talk of extinction, repeated often enough, can crowd out smaller, more immediate questions — whose job this technology takes this year, what it costs to train, who it quietly harms in deployments already happening, not in some future decade. 

Coxon himself is a separate case, and it’s worth saying so plainly. An executive who calls his own product civilisation-ending and keeps building and selling it anyway invites an obvious question: why hasn’t he stopped? Coxon did stop. He gave up his job rather than keep working on something he’d come to believe was genuinely dangerous. Whatever you think of his numbers, that’s a different kind of statement than a warning issued from an office with no plans to close.

The skeptics aren’t wrong that alarm can be good for business. That doesn’t make the alarm false. Both things can be true in the same week’s headlines.

A sharper version of the same doubt comes from inside the industry’s own political circles. David Sacks, an adviser to the current US administration, has gone further than “flex” and called Anthropic’s safety warnings part of a deliberate campaign to saddle smaller competitors with rules the giants can afford and the newcomers can’t — regulation as a moat, dressed up as conscience. Whether or not that charge is fair to any one company, it names a real incentive that sits alongside the sincere one: a company that succeeds in defining the danger also gets a say in who is allowed to build anything at all.

This essay has spent its opening pages inside myths where the danger was real and the warning, however self-interested the messenger, still deserved to be heard. We may not always be able to tell, from where we’re standing, which kind of week we’re living through — the myth’s, or the marketing department’s. That’s not a reason to dismiss either reading. It’s a reason to judge the builders less by what they say about the danger, and more by what they’re willing to give up to contain it. 

Napoleon at the edge of the map

Step out of myth for a moment. History has told this exact story often enough that it stops looking like legend and starts looking like a pattern.

By 1812, Napoleon had every reason to think restraint was for other people. He’d redrawn the map of Europe and beaten Austria and Prussia. So he marched roughly six hundred thousand men into Russia — a campaign his own generals privately doubted — because everything that had happened before told him it would work again.

It didn’t. Russia let the size of his own ambition defeat him. Moscow was burned rather than surrendered. The retreat, starting in October, ran into a winter that killed more soldiers than any battle had. Fewer than one in ten of the men who marched in came home.

Every earlier victory had taught Napoleon one lesson: push further. None of them had taught him when to stop — because none of them had needed to.

This is the part myths only gesture at and history states outright: power that has never met a limit doesn’t learn to look for one. It learns the opposite. It usually takes a disaster of a completely new size to teach the lesson success never could. 

Ashwatthama: a weapon fired in grief

The Mahabharata’s darkest lesson about restraint comes after the war is already, in every practical sense, over.

Ashwatthama’s side has lost. His father is dead. That night, out of grief rather than strategy, he kills the sleeping sons of the winning side in their tents — a massacre the epic never tries to excuse.

Cornered afterward, with nothing left to lose, he does something no disciplined warrior does anywhere else in the story. He fires the Brahmastra — not at an army, not under any of the rules every other user of that weapon follows — straight at the one target that could end the winning family’s bloodline completely: the unborn child in Uttara’s womb.

He’d been trusted with that weapon under the same rules as everyone else. Grief broke every one of them.

The weapon leaves his hand. It can’t be called back — once fired, a Brahmastra travels until it finishes the job, and Ashwatthama was never trained to withdraw one.

Somewhere between the weapon leaving his hand and whatever happens next, an entire bloodline is about to end in a single stroke. Nothing in the story so far tells us anyone is close enough, or fast enough, to stop it.

End of Part One