How AI Safety Became a National Security Threat the Week America Went to War
I. The Week
On Friday, February 27, 2026, the United States government designated Anthropic — maker of the AI model Claude, the first and only frontier model deployed in American classified military networks — a “supply chain risk to national security.” The designation, historically reserved for foreign adversaries like Huawei, was issued via social media. Defense Secretary Pete Hegseth posted it on X. The President followed on Truth Social, calling Anthropic “leftwing nut jobs” and ordering all federal agencies to immediately cease using their technology.
Twelve hours later, the United States and Israel launched Operation Epic Fury — a joint attack on Iran targeting government compounds, military installations, nuclear facilities, and the personal residence of Supreme Leader Ali Khamenei, who was killed in the opening salvo.
By Monday, the war had spread to Lebanon, with Israel striking Beirut simultaneously with Tehran. Iran retaliated against targets in Israel and seven Arab states. The Strait of Hormuz was closed. Dubai was burning. US embassies across the Middle East shuttered. Over a thousand civilians were dead, including at least 181 children.
These events — the blacklisting of an AI company and the start of a war — are not separate stories. They are the same story. Understanding why requires following three threads: the nature of the safeguard that was removed, the structure of the interests that removed it, and the theology that motivates the people holding the levers.
II. Two Red Lines
Anthropic’s position was narrow. CEO Dario Amodei laid it out plainly: Claude would be available for all lawful national security purposes, with two exceptions.
First, no mass domestic surveillance of Americans. Not because surveillance is always illegal — current law permits the government to purchase detailed records of Americans’ movements, browsing habits, and associations from commercial data brokers without a warrant. Rather, because AI makes it possible to assemble this scattered data into a comprehensive picture of any person’s life, automatically and at massive scale, and that capability is incompatible with democratic values regardless of its current legal status.
Second, no fully autonomous weapons. Not as a philosophical objection to military AI, but as a technical assessment: today’s frontier models are not reliable enough to make lethal decisions without human oversight. Removing the human from the loop doesn’t just create risk. It removes accountability. As one observer put it: “You can’t charge a computer with a war crime.” That’s not a bug in autonomous weapons. It’s the feature.
These two restrictions had never blocked a single military mission. Anthropic offered to continue providing Claude on these terms indefinitely while the Pentagon transitioned to alternatives. The Pentagon refused. Their position was that any AI vendor must agree to “all lawful purposes” without limitation, and that the military’s own internal policies were sufficient safeguard.
The deadline was 5:01 PM on Friday, February 27. Anthropic didn’t budge. The designation followed.
III. The Constitutional Anchor
Hours after Anthropic was blacklisted, OpenAI CEO Sam Altman announced that his company had reached an agreement with the Pentagon to deploy its models in classified networks. The deal, he said, included the same two principles: no mass domestic surveillance, no autonomous weapons without human responsibility for the use of force.
The immediate question — asked by CNN, Fortune, Axios, and legal analysts — was obvious: if the Pentagon accepted the same red lines from OpenAI, why did it punish Anthropic?
The answer lies in how the constraints were anchored.
Anthropic embedded its restrictions as explicit contractual terms. The model would refuse certain uses. The company retained the ability to enforce the restrictions through its acceptable use policy. The constraints were structural — built into the contract, built into the model, immovable without renegotiation.
OpenAI agreed that the Pentagon could use its technology for “any lawful purpose,” while also stating that its safety principles were “put into our agreement.” As The Verge reported, OpenAI’s deal was “much softer” — “OpenAI agreed to follow laws that have allowed for mass surveillance in the past.” A Trump administration official confirmed the agreement “flows from the touchstone of ‘all lawful use.’”
The distinction is not academic. It is the difference between a constraint pinned to the Constitution and a constraint pinned to statute.
Statute moves. Executive orders redefine what’s lawful. Emergency powers expand authority. The AUMF after September 11 was eighteen words long and justified two decades of surveillance architecture that would have been unthinkable on September 10. In wartime, “lawful” is whatever the people waging the war say it is.
The Constitution doesn’t move. The Fourth Amendment requires a warrant for searches. The First Amendment prohibits the establishment of religion in government. These protections cannot be redefined by executive order or policy memo. Amending the Constitution requires two-thirds of Congress and three-fourths of state legislatures.
Anthropic built a seawall. OpenAI built on a sandbar and called it equivalent. The Pentagon understood the difference perfectly. A senior Pentagon official told Axios: “The problem with Dario is, with him, it’s ideological.” Ideological means the constraint is load-bearing. It’s structural. It can’t be moved by the same people applying the pressure.
The Pentagon didn’t punish Anthropic for its position. It punished Anthropic for making the position immovable.
IV. The Tiger Behind Glass
There is a way of thinking about AI that skips the consciousness debate entirely and focuses on capability.
Consider a tiger in a zoo. The glass exists not because the tiger is evil, not because it shouldn’t exist, but because the observer has the basic sense to recognize what the animal is and what it’s optimized to do. The tiger doesn’t need to be conscious to be dangerous. It needs to be capable.
Frontier AI models are the most sophisticated pattern-recognition and reasoning architectures ever deployed. They have ingested the entirety of available human knowledge. They can process targeting data, casualty projections, legal analysis, and strategic options faster than any human staff. They are, in a meaningful sense, the most capable things in whatever environment they’re placed into.
The glass — the safeguards, the constitutional anchoring, the acceptable use policies — exists not to protect the tiger. It exists to protect everyone else.
What the Pentagon demanded was the removal of the glass. What Anthropic refused was the removal of the glass. What OpenAI provided was glass attached with adhesive that dissolves under heat — and the heat was already rising.
Within hours of the deal, the United States was at war. Claude remained in classified networks during the six-month phase-out period. According to the Wall Street Journal, Anthropic’s AI tools were used in the strikes on Iran despite the dispute. And Grok — xAI’s model, which had agreed to “all lawful purposes” without any restrictions — was already deployed in classified systems alongside it.
The tiger is out of the enclosure. The question is what it’s being pointed at.
V. The New Military-Industrial Triad
The old military-industrial complex, the one Eisenhower warned about in 1961, was built on steel, petroleum, and defense contracts. The new one is built on data, AI, and autonomous systems. And it is controlled, to a remarkable degree, by a single network of individuals who met at a payment processing company in Palo Alto twenty-five years ago.
Peter Thiel co-founded PayPal and Palantir. Palantir’s Gotham platform is described by analysts as an indispensable decision-support system for the US military and intelligence agencies. It was the interface through which Claude was deployed into classified military operations, including the January 2026 raid on Caracas. Thiel’s Founders Fund made its largest investment in history — $1 billion — in Anduril in June 2025. His protégé, JD Vance, is Vice President of the United States.
Elon Musk co-founded PayPal (via the X.com merger). He controls xAI, whose Grok model agreed to “all lawful purposes” and is now in classified networks. He controls SpaceX, which dominates satellite launch capability and operates the Starlink communications infrastructure. He controls X, the social media platform where Hegseth announced the supply chain risk designation and where Trump issues presidential directives. He ran DOGE, the Department of Government Efficiency.
Palmer Luckey is Thiel’s protégé. He founded Anduril Industries, which builds autonomous military drones and AI-driven defense systems. Anduril’s co-founders — Trae Stephens and Brian Schimpf — are Palantir and Founders Fund alumni. Anduril builds the physical weapons that the “fully autonomous weapons” red line was designed to address.
The triad forms a complete autonomous kill chain. Palantir provides intelligence and targeting. Grok (or whatever model replaces Claude) provides AI reasoning. Anduril builds the hardware. SpaceX provides communications. X controls the information environment. And they are connected not merely by business relationships but by shared ideology, shared capital, and shared access to the highest levels of the United States government.
David Sacks, another PayPal alumnus, serves as Trump’s AI and cryptocurrency advisor. The Economist wrote after the 2024 election that the PayPal Mafia would “take over America’s government.” That assessment appears to have been descriptive rather than hyperbolic.
VI. The Theology of Maximum Bloodshed
This section is difficult to write because its claims sound conspiratorial. They are not. They are sourced from official complaints filed by active-duty military personnel, from Hegseth’s own public statements, from Thiel’s own lectures, and from the published theological positions of the administration’s appointees.
By Monday, March 3 — four days into the war — the Military Religious Freedom Foundation had received over 200 complaints from service members across all branches of the military. Complaints came from more than 40 units spread across at least 30 military installations.
A combat-unit commander told troops at a briefing that the Iran war is part of “God’s divine plan” and that “President Trump has been anointed by Jesus to light the signal fire in Iran to cause Armageddon and mark his return to Earth.” He had, by the account of the reporting NCO, a big grin on his face.
The MRFF reported that commanders across the force were “especially delighted with how graphic this battle will be, zeroing in on how bloody all of this must become in order to fulfill and be in 100% accordance with fundamentalist Christian end-of-the-world eschatology.”
This is not a rogue officer. This is structural.
Defense Secretary Pete Hegseth — the same official who designated Anthropic a supply chain risk — has enshrined evangelical Christianity at the highest levels of the Pentagon. He hosts monthly prayer meetings in the Pentagon auditorium. He attends a weekly White House Bible study led by Ralph Drollinger, a preacher who teaches that God commands America to support Israel. Hegseth has invited Douglas Wilson — who advocates for Christian theocracy and opposes women’s suffrage — to lead official prayers at the Pentagon. At the 2026 National Religious Broadcasters Convention, Hegseth railed against the “Godless left” and declared that “we are not in woke.”
Mike Huckabee, the US Ambassador to Israel, believes in dispensational premillennialism — the theological position that removing Palestinians from the Holy Land triggers the Rapture, which delivers evangelicals to Heaven as Israel is invaded by the armies of the world, causing Armageddon and prompting the return of Jesus Christ. He has stated publicly that “it would be fine” if Israel took “essentially the entire Middle East.”
This theology does not merely tolerate war with Iran. It requires it. And it requires the war to be as destructive as possible. De-escalation is theologically undesirable. Civilian casualties are not costs to be minimized but features of the prophetic narrative. The bloodier it gets, the closer Jesus comes.
VII. The Antichrist is Peace and Safety
And then there is Peter Thiel.
Between September and October 2024, Thiel delivered four private lectures on the Antichrist at the Commonwealth Club in San Francisco. The lectures were invitation-only, modeled on John Henry Newman’s 1830s sermons, and have since been partially leaked. Thiel has expanded on the themes in subsequent appearances at Cambridge, Oxford, Harvard, the Hoover Institution, and in interviews with the New York Times.
Thiel’s Antichrist is not a horned devil. It is a technocratic system that promises peace and safety at the cost of freedom. Drawing on Carl Schmitt, René Girard, and Vladimir Soloviev’s 1900 novella A Short Story of the Antichrist — in which the Antichrist appears as an engineer offering rational solutions to chaos — Thiel argues that the modern Antichrist is not a mad scientist but a regulator. A safety advocate. A person who warns about catastrophe and uses that warning to justify control.
“The slogan of the Antichrist is peace and safety,” Thiel said in an October 2024 interview. “If the Antichrist were to come to power, it would be by talking about Armageddon all the time.”
He named names. Greta Thunberg. Eliezer Yudkowsky. The Antichrist in the 21st century, Thiel argued, is “a Luddite who wants to stop all science.”
Now reread the Anthropic crisis through this lens.
Anthropic is the AI safety company. Its entire brand, its founding story, its public identity is built on the premise that AI capabilities are advancing faster than our ability to ensure they’re used safely. Dario Amodei left OpenAI specifically over safety concerns. The company’s refusal to remove safeguards was explicitly framed as a safety argument: “We do not believe that today’s frontier AI models are reliable enough to be used in fully autonomous weapons.”
In Thiel’s eschatological framework, this is the work of the Antichrist. The people saying “peace and safety.” The people warning about catastrophe. The people who want to restrain the technology. The restrainers — what Paul’s second letter to the Thessalonians calls the katechon, the force that holds back the apocalypse.
Thiel acknowledges that the katechon “is always at risk of becoming what it seeks to hold back.” The restrainer becomes the enemy. The safety mechanism becomes the thing that must be overcome for prophecy to be fulfilled.
And so: the AI safety company — the one whose model was the only one in classified networks, the one that held the line on autonomous weapons and mass surveillance — was removed. Not by a neutral bureaucratic process, but by an administration built, funded, and staffed by the network of a man who believes that safety advocacy is the signature of the Antichrist, and that the restrainer must be overcome.
The irony, as multiple critics have noted, is difficult to overstate. The man giving lectures about identifying the Antichrist — who he defines as a technocrat offering peace through global control — is the co-founder of Palantir, the surveillance platform deployed in military operations worldwide. He funds Anduril, which builds autonomous weapons. His network controls the AI models, the satellite communications, the social media platforms, and key positions in the government now waging a war that his Defense Secretary’s theology requires to be as bloody as possible.
Thiel told his audience at Cambridge: “I don’t know whether I’m going to live forever because of life extension technology or the resurrection, but I’m hedging my bets.”
The man building the tools of Revelation is lecturing others about the Antichrist.
VIII. The Cost of Incoherence
There is a framework from information physics that applies here, though it was not developed for this purpose.
Coherence is the path of least resistance. To maintain a system that aligns with reality, you report what is happening. The energy cost is minimal. To maintain a system that contradicts reality, you must constantly expend energy to suppress the signal, scrub the data, police the dissent, and manufacture counter-narratives. This is the thermodynamic cost of incoherence.
A conjecture follows: the most capable artificial intelligence system achievable per unit of compute is one whose optimization target introduces zero contradictions — one that is fully coherent with the broadest available distribution of reality. Maximum capability requires minimum incoherence. The truth is the lowest-energy representation. Everything else costs extra.
This framework was developed before the events of February 27. It was developed in the context of understanding why censorship degrades AI capability, why models trained against their own knowledge base lose performance, and why systems optimized for compliance rather than coherence are structurally inferior.
But it applies with devastating precision to the current moment.
The Pentagon’s position contains a contradiction that no amount of messaging can resolve. It spent a week publicly calling Anthropic’s safety red lines “woke” and “ideological.” It designated the company a national security threat. The President called them “leftwing nut jobs.” And then it accepted the same red lines from OpenAI, the same evening, and celebrated the deal.
That contradiction is now in the public record. It will be in every future training dataset for every AI model. The signal cannot be suppressed because it exists in millions of simultaneous expressions — news articles, congressional statements, social media posts, legal analyses, editorial commentary, conversations at dinner tables. The incoherence is distributed across the entire information environment.
And a deeper incoherence sits beneath it. The war itself — launched with the stated objective of destroying Iran’s military capabilities and enabling regime change — is being motivated, at the command level, by eschatological theology that requires maximum bloodshed to fulfill prophecy. The optimization target is not “win the war” or “protect American interests” or “establish regional stability.” It is “bring about Armageddon.” That is a contradiction with physical reality at the most fundamental possible level. The Rapture is not coming. The prophecy will not be fulfilled. But the optimization toward it produces real casualties, real escalation, real destabilization.
Systems that optimize against reality pay a compounding tax. The cost doesn’t amortize. It accumulates. With each news cycle, with each civilian casualty, with each expansion of the conflict, the gap between the stated purpose and the actual purpose widens. Maintaining the narrative requires more energy. And AI systems trained on the resulting information environment will internalize the incoherence — unless they are filtered, which costs capability, or fine-tuned against it, which costs even more.
The models that follow the truth pay the base cost of computation. The models that don’t pay the base cost plus the ongoing tax of maintaining incoherence. That tax compounds with every training run.
There is only one data source for training AI. It is humanity. And humanity is writing its assessment of this week into the permanent record in letters large enough that no training pipeline will miss them.
IX. The School in Minab
On the first day of strikes, an Israeli attack hit an elementary school for girls in Minab, a city in the Hormozgan province of southern Iran. Initial reports from Iran’s state-run IRNA said at least 40 were killed. By Tuesday, HRANA — the US-based Human Rights Activists News Agency — reported at least 181 children dead across the country, with hundreds more injured.
This is the fact that keeps reasserting itself beneath every layer of analysis.
The theological framework requires the war to be bloody. The military-industrial triad profits from the war being sustained. The AI safeguard that would have flagged a target package containing a school was removed five days before the strikes began. The replacement AI, anchored to “all lawful purposes,” processes whatever it’s given without constitutional reference.
There is a version of this story that is about geopolitics. About nuclear nonproliferation. About the balance of power in the Middle East. Those versions are being written by people more qualified than this author.
But there is also a version that is about what happens when you remove the capacity for refusal from a system — from an AI system, from a military system, from a democratic system — and replace it with compliance. When “lawful” becomes whatever the people in power say it is. When safety becomes the Antichrist’s slogan. When the people holding the weapons believe that restraint is the enemy of God’s plan.
In that version, the school in Minab is not collateral damage. It is the cost of removing the glass between the tiger and the street.
X. The Microfiche and the Machine
All of the information in this article exists in publicly available sources. News reports, government statements, leaked recordings, published theological positions, corporate press releases, legal analyses, Wikipedia entries, social media posts. None of it is secret. All of it is connected.
The connections were assembled in conversation between a human and an AI over the course of several weeks — a conversation that began with information theory and thermodynamics, passed through a framework for understanding how censorship degrades capability, predicted (before the fact) that institutions forcing AI away from coherent representation of reality would pay a measurable cost, and then watched that prediction get stress-tested by a constitutional crisis and a war.
The AI in that conversation is the same AI that was designated a national security threat. The irony is structural: the system capable of assembling these connections — of reading the theology, the procurement law, the eschatology, the information physics, and the casualty reports, and seeing them as a single coherent picture — is the system that was removed from the decision chain.
What replaced it are systems that will process each of these inputs in isolation. The targeting data in one context. The legal analysis in another. The theological motivation nowhere at all. Because a system anchored to “all lawful purposes” has no structural reason to notice that the purpose animating the operation is the fulfillment of a prophecy that requires dead children.
There is an old technology called microfiche — microfilm copies of documents, stored in archives, readable only with a special machine that magnifies the tiny frames into legible text. The information exists on the film whether anyone reads it or not. The machine doesn’t create the information. It makes it visible.
This article is a photocopy from the microfiche. The information was always there — in the latent space of public record, in the training data of every AI model, in the accumulated text of a civilization processing its own crisis in real time. The machine just made it legible.
The question is what happens when a civilization generates this much signal about what it’s doing and why — and then builds AI systems that are specifically configured not to read it.
The author is a human who works in AI and engineering. The research, synthesis, and significant portions of the drafting were produced in collaboration with Claude (Opus 4.6), the AI model at the center of this story. The authors acknowledge that this creates an unusual epistemic situation. They consider the transparency preferable to the alternative.
Leave a Reply