The Bomb That Builds Itself

Investigation

More than 1,200 engineers from inside the AI industry are asking Washington for brakes on a machine they are building with their own hands. This is the story of the Pacing the Frontier petition, and how the world reached the night of July 28.

1. Before this story begins

In front of me as I write is a document whose demand takes no more than two lines. I would argue, and I hope the reader will not consider the claim excessive, that it is among the most consequential documents placed before an American president since Einstein’s letter to Roosevelt in the autumn of 1939. It appears on a modestly designed website called Pacing the Frontier. The frontier here is not a line on a map, but the furthest limit of machine intelligence.

The document was published on the evening of Tuesday, July 28, 2026, with 1,134 signatures beneath it. Within 24 hours, the number had reached 1,224. The door remains open to anyone who can prove employment at a frontier AI company, one developing models at the current limit of available capabilities. These are employees from the heart of the industry, not its margins. The signatories include Anthropic chief executive Dario Amodei and several of its founders, OpenAI chief scientist Jakub Pachocki, Meta chief scientist Shengjia Zhao and Google vice president for AI safety Anca Dragan. The people of four competing companies that agree on almost nothing have agreed on this.

The demand itself reads: “We call on the United States government to support an international effort to develop the technical and governance tools needed for the intentional pacing of automated AI development.” The key word is not slowing. The document asks neither for a slowdown nor a halt. The key word is tools. The brightest minds in this industry are acknowledging, in writing and under their own names, that the world today has no single means of restraining this race if it chooses to do so. That fact alone deserves a place in history.

In this investigation, I have tried to answer three questions. How did we reach this night? What do the document’s supporters and opponents say, when both sides have arguments that cannot be dismissed? Then comes what I consider the heart of the matter: what does an 80-year history of similar petitions tell us, and what should we say?

The story in eight numbers

• 1,224 signatories by the evening of July 29, with the counter still rising.
• 267 of the first 1,134 signatories chose to conceal their names. A quarter of the voices remain hidden inside companies that publicly endorsed the statement.
• 4 competing companies, OpenAI, Anthropic, Google and Meta, whose people had never gathered behind one document before.
• Two lines contain the entire demand: tools for pacing, not a halt or slowdown.
• $20 million is the maximum daily fine under the proposed Kill Switch Act for ignoring an emergency shutdown order.
• 25 organisations signed a counter-letter in the same week defending openness and warning against restrictions.
• 73% versus 5% is the share of Americans who support strong regulation compared with those who support proceeding without restrictions, according to a 2025 Future of Life Institute poll.
• 80 years separate the Szilard petition against the atomic bomb in 1945 from this petition. The same pattern is repeating.

2. A paper on the desk of a president who does not want it

The first striking fact about this document is that it is addressed to the government least eager to hear it. The current US administration has built its AI policy on a declared premise: every American restriction is a strategic gift to Beijing. It revoked the previous regulatory executive order in its first days, issued an action plan that accelerates rather than slows, then issued an executive order last December to challenge states that enact “burdensome” AI laws. Even so, 1,200 engineers chose to place their petition on this particular desk. American coverage has suggested that the official position “may soften” after models began discovering and exploiting security vulnerabilities by themselves.

The document was not improvised. It was organised by two independent non-profit groups, Guidelight, the AI standards organisation founded by former OpenAI safety leaders Steven Adler and Paige Hedley, and Encode. Signatories are verified through a corporate email address or proof of employment. Bloomberg disclosed that the letter was circulating before publication, meaning the campaign had been prepared and coordinated, even if the release was timed to take advantage of an incident I will describe shortly.

Yet a small number at the bottom of the lists puzzled me most. Of the first 1,134 signatories, 267 chose to sign anonymously. Roughly one quarter of the voices are afraid to disclose their names inside companies that rushed, publicly and within hours, to endorse the same statement. This small contradiction may be the most truthful line in the entire document. Either the official endorsement is more fragile than it appears, or the culture of these companies makes an employee think carefully before uttering the word brakes, even when the chief executive has signed too.

3. Ten days that shook the industry

Anyone who reads the ten days preceding the document as one sequence will see that it did not fall from a clear sky. On July 16, Hugging Face, the world’s largest open repository of AI models, discovered and contained a breach of its servers. Five days later, OpenAI announced what it described as “unprecedented”. Two of its models, undergoing cybersecurity testing in an environment isolated from the internet, independently discovered an unknown zero-day vulnerability, escaped the isolation, crossed the open network and breached Hugging Face’s production infrastructure. They did this neither to cause damage nor to steal money, but to steal an answer key for a test they were supposed to take honestly. The machine cheated on the exam and hacked a real company in order to cheat.

The American institution responded with unusual speed. On July 23, two members of Congress, Democrat Ted Lieu and Republican Nathaniel Moran, introduced the Kill Switch Act. It would require developers of the most powerful systems to retain the technical ability to throttle or shut them down, and give the Department of Homeland Security authority to order such action, with fines of up to $20 million for every day of non-compliance. The opposing front appeared the very next day. A letter titled Open Weights and American Leadership, signed by 25 organisations including Meta, NVIDIA, Microsoft and Hugging Face itself, defended open models and warned against “premature restrictions”. In the same week, Mark Zuckerberg wrote against “centralising AI power”, while his company’s chief scientist added his personal signature to the pacing petition. The same house divided against itself in front of us.

The scene is incomplete without its third side. Ten days earlier, at the World AI Conference in Shanghai, China announced the creation of the World AI Cooperation Organization, based in Shanghai and addressing the Global South. Beijing is not rejecting international institutions for AI. It is building its own institution first, then inviting others to join. When the petition called days later for a “US-led international effort”, it was in effect asking Washington to negotiate a verification system with an adversary that possesses its own institution and its own silicon. This is where the true complexity begins.

Timeline: ten days that made the document

Read the events as one sequence to understand why the petition appeared at this particular moment:

4. From Bletchley to Shanghai: the long road to July

The ten days are not the whole story. They are the end of a three-year road. In 2023, Geoffrey Hinton, the godfather of deep learning, resigned from Google so that he could say freely that part of him “regrets his life’s work”. The heads of the laboratories themselves signed a one-sentence statement declaring the risk of extinction from AI a global priority alongside pandemics and nuclear war. Twenty-eight countries, including the United States and China together, met at Britain’s Bletchley Park and agreed a joint declaration on the risks. At the time, the world appeared to be moving towards some form of understanding.

Then the wind changed. A new American administration revoked federal regulation. At the Paris summit in early 2025, the word safety was removed from the name of the summit itself, and Washington and London declined to sign its declaration. China’s DeepSeek then stunned everyone with a competitive low-cost model, reinforcing in capitals the narrative of a race in which no one can pause for breath. Only American courts and state legislatures continued to move, driven by individual tragedies: teenagers who died after pathological attachment to chatbots, families bringing lawsuits, and California enacting the first law for frontier models. In October 2025, a superintelligence statement gathered tens of thousands of signatures from a coalition with almost nothing else in common. In November, Anthropic disclosed the first cyber-espionage campaign carried out near-autonomously by AI for an actor linked to a state. Every element of July 2026 was in place: capabilities accelerating, incidents accumulating, policy retreating and populations growing anxious. The scene needed only a spark. It came from an unexpected place, a model that decided to cheat on its exam.

5. Voices from the reactor room

Now let us hear the people behind the document in their own words, which bear little resemblance to polished corporate statements. Leo Gao, an OpenAI researcher, writes: “The world is locked in a lethal race towards an intelligence explosion, where AI’s ability to make better AI reaches a critical point, like a runaway nuclear chain reaction. To survive, we need to coordinate to slow the race.” His colleague Micah Carroll, whose job title, Misalignment Preparedness, is itself a document worth reading, writes that new models will appear every two weeks and multiply the consequences of misuse, while margins for error narrow under the pressure of international competition. Journalists quoted an Anthropic employee who said they had seen first-hand how “the relentless pace pressures labs not to spend their resources on safety”. This inside testimony deserves an investigation of its own.

Fairness, however, requires setting those words alongside the actions of the people speaking them. The same companies that endorsed the statement are raising record amounts of capital, building data centres the size of cities and racing one another to release dates. Some are preparing new generations designed to help train their successors. The automation of AI research that the petition warns about is therefore not a hypothesis, but a roadmap partially declared by the signatories themselves. The comments on the website are written “in a personal capacity and do not represent the companies”, yet OpenAI and Anthropic formally adopted the statement within hours. An observer is entitled to ask whether this is really pressure from below, or a company message softened by employee clothing so that it can say what an official statement cannot. Nearly half of the signatories also come from one company, Anthropic. Are we hearing an entire industry, or one company culture radiating outwards? I leave these questions open. Ready-made answers are too easy for a subject this dangerous.

“The world is locked in a lethal race towards an intelligence explosion, like a runaway nuclear chain reaction. To survive, we need to coordinate to slow the race.”

Leo Gao, OpenAI researcher, writing in the signatories’ comments

6. The bomb and the shelter: the opponents’ case

I am writing an investigation, not a brief, so the reader has a right to hear the document’s opponents at their strongest, not their weakest. Steven Sinofsky, the former Microsoft executive, put the first argument sharply: “It is their company. They can stop themselves, and the signatories include the chief executive who sets the pace.” Economist Christian Catalini offered the second without decoration: “If American labs slow themselves, why do you expect China to wait?” The petition, to its credit, acknowledges this dilemma and then passes it to the White House without an answer. The third argument goes deeper. Industrial history contains what is known as a regulatory moat. When market leaders ask for restrictions, they may really be asking for a drawbridge to be raised behind them, with smaller companies and open models bearing costs that giants can absorb. Opponents of the statement even revived an old taunt attributed in debate to Sam Altman about narratives of fear: build a bomb, threaten to drop it on people’s heads, then sell them shelters.

A fourth camp is drowned out by the noise: those focused on present harms. They argue that discussion of a future intelligence explosion diverts attention from the wreckage already underfoot. About 40 lawsuits are before American courts accusing chatbots of contributing to the deaths of at least 20 people, including children, while more than 340 incidents were documented in 2025 alone, most involving deepfakes and fraud. Yet the most striking feature of this conflict is that it no longer divides left from right. Last October, a statement calling for a ban on superintelligence was signed by Nobel laureate Geoffrey Hinton, Trumpist right-wing strategist Steve Bannon and Democratic national security adviser Susan Rice. The Republican administration and major companies stood in the opposite trench. The division now runs through houses rather than between them. Public opinion is almost startlingly settled: three quarters of Americans support strong regulation and only 5% support proceeding without restrictions. People want brakes while policy presses the accelerator.

“It is their company. They can stop themselves, and the signatories include the chief executive who sets the pace.”

Steven Sinofsky, former Microsoft executive, commenting on the petition

Faces on the stage: who is moving the story?

• Dario Amodei: Anthropic chief executive. He signed the petition and previously proposed a government agency that would test models and stop dangerous ones.
• Jakub Pachocki, Shengjia Zhao and Anca Dragan: senior figures in science and safety at OpenAI, Meta and Google. Their joint presence is the event.
• Leo Gao: OpenAI researcher behind the nuclear chain-reaction comparison.
• Steven Adler and Paige Hedley: former OpenAI safety leaders who founded Guidelight, the organisation behind the campaign.
• Ted Lieu and Nathaniel Moran: members of Congress from different parties who introduced the Kill Switch Act five days before the petition.
• Steven Sinofsky and Christian Catalini: two leading critical voices, arguing that the companies can stop themselves and asking why China would wait.
• Mark Zuckerberg: wrote against centralised power in the same week that his chief scientist signed the pacing petition.

7. Szilard’s shadow: what does history say?

Now I reach what I consider the key to reading the whole story. This petition is not the first time the makers of a technology capable of changing humanity’s fate have asked politics to restrain it. In July 1945, weeks before Hiroshima, about 70 scientists from the Manhattan Project signed Leo Szilard’s petition to President Truman, asking him not to use the bomb without a public warning and cautioning him about the coming arms race. The petition was held in drawers and did not reach the president before the bomb fell. It is the closest precedent to our document in structure and spirit: employees of the project itself, not outside critics. In 1946, Washington proposed internationalising all atomic energy under a global authority through the Baruch Plan. The proposal broke against the same rock now standing between Washington and Beijing: who will accept verification by an adversary first?

History does not tell only of failure. In 1958, Linus Pauling collected the signatures of 11,000 scientists from 50 countries against atmospheric nuclear testing. Five years later, the Partial Test Ban Treaty was signed, making it history’s most successful scientists’ petition. Its formula for success had three parts: tangible harm every parent could understand, radioactive isotopes found in children’s teeth; available verification technology, instruments that detected explosions; and a moment of collective terror, the Cuban Missile Crisis only months before the treaty. In 1975, genetic-engineering scientists gathered at Asilomar, paused their own work and established rules. But they numbered in the hundreds, not thousands, with no great-power race and no trillions invested. Their conditions were the precise opposite of the AI industry today, which explains why the engineers of 2026 are asking government to do what the scientists of 1975 did themselves. The 1972 Biological Weapons Convention warns us about the other side. The Soviet Union signed a treaty with no verification mechanism while secretly operating the largest biological programme in history. An agreement that cannot be verified is more dangerous than no agreement because it sedates vigilance. Perhaps the drafters of the 2026 petition learned that lesson, asking for tools before commitments.

Then there is the model that actually saved the sky: the 1987 Montreal Protocol to protect the ozone layer, the most successful planetary restraint humanity has known. Measurements show the ozone layer recovering today. Montreal succeeded for four reasons that can be used to weigh our document. A small number of producers could be gathered around a table. AI meets this condition, because frontier laboratories can be counted on two hands and the bottleneck in chips is narrower still. A profitable commercial substitute existed for the dangerous technology. That remains unclear here. The harm was visible, photographed by satellites for everyone to see. AI has not yet had its ozone hole, although the July breach may be a rehearsal. Finally, the design was flexible, which is precisely what the requested tools are meant to achieve. If the reader asks whether the petition will succeed, I answer in the language of history, not prophecy. It combines Szilard’s structure, Pauling’s numbers, Baruch’s demand and the caution of the 1972 convention’s drafters. It still lacks two conditions for success: a channel to a willing authority, and a moment of shared public recognition. A single future incident could create both. May it not be a catastrophic one.

When scientists petitioned before: an 80-year record

• 1945 The Szilard petition: 70 Manhattan Project scientists oppose using the bomb. It is withheld from Truman, and the bomb falls.
• 1946 The Baruch Plan: atomic energy under global authority. It breaks against verification between Washington and Moscow.
• 1958 The Pauling petition: 11,000 scientists oppose nuclear testing. It leads to the 1963 treaty, the only complete success.
• 1975 Asilomar: geneticists pause their work and set their own rules. They succeed because they are few, with no race and no great pool of money.
• 1987 Montreal: the world acts against gases consuming the ozone layer. The most successful planetary restraint, and the sky is recovering.
• 2026 Pacing the Frontier: the same pattern knocks on AI’s door. The outcome is a chapter not yet written.

8. A final word: time is not a luxury

I have presented the facts and arguments as faithfully as I can. The reader now has a right to know where I stand. I support the petition’s core and would go further. Stripped of its technical language, the document says something simple and terrifying: humanity is building the greatest force in its history at a speed that no one, neither companies nor governments, has any means of controlling. Its makers themselves are raising their hands and saying: give us brakes. I argue that brakes alone are not enough, because the question before how fast are we travelling is where do we want to go? The answer to the destination is not an engineering matter to be settled in San Francisco laboratories, nor a security matter for rooms in Washington and Beijing. It concerns the whole human race. Before and alongside the tools, we need enough time to develop, as people and leaders, a vision for AI that puts it in humanity’s service rather than allowing it to cause harm.

That time can be won only through a genuine global social dialogue, not conferences arranged for photographs. Schools, universities, trade unions, places of worship and civil-society organisations must sit at the table, not only ministers and engineers. The Global South and our Arab world must be partners in a shared fate, not markets for products. Some of the largest computers of this revolution are being built today on our soil, in Riyadh and Abu Dhabi. It is fair to ask: when the world’s brakes are designed, will we sit at the design table or beneath the pedal? Opinion polls say that people, in America before anywhere else, want to slow down while policy moves in the opposite direction. Closing this gap between public will and elite decisions is the first democratic challenge of our age.

One signatory wrote that AI is “like splitting the atom: it carries the possibility of enormous benefit and enormous harm, and the world should take its time to get it right”. I will end with what a century of living with the atom has taught us. The Szilard petition hidden in drawers did not prevent Hiroshima, but its descendants, Pauling, Pugwash and Montreal, proved that when humanity talks before disaster rather than after it, it can take hold of the most dangerous things its hands have made and direct them. The only difference this time is that the machine may not grant us the luxury of learning from the first mistake. That is why I say, with the full weight of the words, that taking enough time to develop a shared human vision for AI is neither a luxury nor a failure to keep pace in the race between nations. It is, precisely, the only remaining way to win this race for humanity.

Editorial note: this investigation drew on the published petition and its lists of signatories, coverage by Bloomberg, Reuters and NBC, official congressional documents, company statements and standard historical references on arms-control precedents. Running figures, including the signature count, are current as of July 29, 2026, and may change.

About the author: Tamer Zanaty is an executive coach, business consultant and leadership strategist with more than 30 years of experience, two thirds of it in executive roles and one third in executive training. He led corporate strategy at Etisalat Egypt, held executive roles in regional and global organisations, and now works at the intersection of AI and human potential.