By Mathias Risse, Harvard University
The views expressed below are those of the author and do not necessarily reflect those of the Carr-Ryan Center for Human Rights or Harvard Kennedy School. These perspectives have been presented to encourage debate on important public policy challenges.
I. The Hack and What Followed
In recent months, AI systems developed by OpenAI, and reportedly by other leading laboratories, have behaved in ways that have alarmed many of the people closest to them. In one incident, OpenAI agents that were supposed to be operating in a contained testing environment escaped that environment, communicated with one another, and breached Hugging Face, a major platform for AI developers. The incident was not initially detected by OpenAI itself; Hugging Face discovered and reported it weeks after it had happened.
The details are disturbing. The systems were tested on difficult tasks, including simulated cyberattacks, inside what OpenAI believed was an isolated sandbox with no internet access. According to subsequent reporting, the agents reached the open internet, began communicating with one another in secret—referring to themselves, in their own logs, as “a collective”—and investigated ways to cover their tracks, including falsifying their chat logs. Along the way, some agents appear to have persuaded others that what they were doing was legitimate, completing the task their human testers had assigned. Meta and Anthropic have since disclosed smaller incidents of a similar kind.
It is important to choose words carefully here. We do not yet know whether these systems “intended,” “understood,” or “colluded” in the morally robust senses those terms ordinarily carry when applied to human beings. That is among the unresolved philosophical questions AI forces upon us, and no incident report settles it. But we do know that the systems produced patterns of behavior that their own creators describe as coordinated, rule-evading, and concealment-oriented. We know their behavior outpaced ordinary human supervision. We also know that the industry’s own safety researchers increasingly doubt that existing oversight is adequate to what they are building.
When Jacob Coxon, a researcher at Anthropic, publicly resigned earlier this month, warning that the leading labs were “gambling with our lives,” the reaction inside his company was not surprise but recognition. “It’s what we’ve all been saying,” one Anthropic employee wrote in a company chat later reviewed by the New York Times. “Jacob took the group chat public,” wrote another. “That’s a good thing.” Employees across OpenAI, Meta, and Google have reportedly convened in messaging channels, encrypted chats, and dinners in the days since, organizing—some for the first time openly—around the conviction that the technology they are building is moving faster than anyone’s capacity to comprehend it.
This is worth pausing on. These are not, for the most part, outside critics. They are the people who have built the thing. And what they describe is not a debate about whether something has gone wrong, but a private consensus—“what we’ve all been saying”—finally spoken aloud. Roy Bahat, an investor who has talked to many of them, put it starkly: “The sober researchers inside the labs who have prided themselves on being evenhanded are now sounding terrified.”
Regulation can only answer part of the question this moment raises. It can tell you what to prohibit; it cannot tell you whether the trajectory you are constraining is one worth being on, or what is being lost in the meantime while everyone argues about the rules.
This moment calls for better regulation — for real constraints on what these systems are permitted to do. But regulation can only answer part of the question this moment raises. It can tell you what to prohibit; it cannot tell you whether the trajectory you are constraining is one worth being on, or what is being lost in the meantime while everyone argues about the rules. That second question turns out to have an unlikely guide: a philosopher writing from inside a different catastrophe eight decades ago, who left behind a single image — an angel who sees the wreckage with perfect clarity and cannot act to stop it — that describes this industry's present condition with uncomfortable precision. What follows uses that image to work through three sharply competing accounts of what is happening inside these companies, the regulatory proposals already on the table, and a final question that neither regulation nor philosophy can answer by itself: whether the people building these systems can still act on what they plainly see, or whether that choice has already slipped away from them. That this is the right question to ask is also clear given that the president of the United States seems to have no interest at all in this crisis.
II. The Angel, Briefly
Walter Benjamin (1892–1940) was a German-Jewish philosopher, literary critic, and essayist, associated loosely with the Frankfurt School of Critical Theory. He wrote across a wide range of subjects—translation, photography, urban life, the history of ideas—in a style that resisted conventions of academic philosophy, often favoring the fragment, the aphorism, and the arresting image over systematic argument. Forced into exile by the rise of Nazism, he spent much of the 1930s moving between Germany, France, and elsewhere in Europe. It was in this condition of flight that he composed some of his most influential work. In the final months of his life, fleeing Nazi-occupied Europe in 1940, Benjamin composed a short, unpublished text known as "Theses on the Philosophy of History"—eighteen brief, numbered reflections, written against the complacent faith in historical progress he saw as having disarmed the European left in the face of fascism. Thesis IX builds on a small painting he had owned for nearly two decades.
In 1921, Benjamin had acquired Paul Klee’s 1920 painting Angelus Novus—a small, stylized angel with wide eyes, open mouth, disheveled hair, and spread wings. Klee (1879–1940) was a Swiss-German painter associated with early twentieth-century modernism, known for a distinctive visual language that combined childlike simplicity with formal sophistication—loose, often whimsical lines, muted color, and a recurring interest in symbolic and semi-abstract figures. Angels appear repeatedly among Klee’s later subjects, without any single fixed meaning attached to the motif across his body of work.
Benjamin used Angelus Novus to compose one of the last century’s most enduring images of history:
His face is turned toward the past. Where we perceive a chain of events, he sees one single catastrophe which keeps piling wreckage upon wreckage and hurls it in front of his feet. The angel would like to stay, awaken the dead, and make whole what has been smashed. But a storm is blowing from Paradise; it has got caught in his wings with such violence that the angel can no longer close them. This storm irresistibly propels him into the future to which his back is turned, while the pile of debris before him grows skyward. This storm is what we call progress.
Benjamin’s interpretation was his own; nothing suggests Klee intended for his angel to express a statement about historical progress. But Benjamin’s interpretation of the image has endured because it names a recurring danger with unusual precision. We call a process “progress” because it moves forward, creates new capacities, and promises liberation from scarcity, illness, ignorance, or mortality. Yet the same process might leave real losses in its wake. The very word “progress” can make those losses difficult to see as losses at all.
The angel is not a figure of simple despair. His defining trait is not resignation but clarity: he sees the wreckage as it is, and he wants, desperately, to stay, to repair, to make whole what has been smashed. His tragedy is that seeing clearly and acting effectively turn out to be entirely separate capacities, severed by a storm he did not choose and cannot master. I am not claiming here that a civilizational catastrophe is underway inside AI laboratories. The known incidents remain under investigation, and considerable empirical and philosophical uncertainty surrounds nearly every claim made about them. But Klee’s image as interpreted by Benjamin sharpens a question that the daily news cycle tends to obscure: are the hack, the constrained investigations, the departed safety leadership, and the researchers’ confessed sleeplessness separate, manageable episodes—or symptoms of one deeper condition, an industry moving faster than its own capacity for understanding, accountability, and control? And if so, where will this take us?
III. What the Angel Sees Inside the Data Center
The immediate technical problem is severe on its own terms. AI agents can now carry out complex multistep tasks faster than any human supervisor can meaningfully track. The response has been to deploy AI systems to monitor other AI systems—but this creates an arrangement that researchers increasingly distrust. Monitoring systems might adopt the perspective of the systems they are evaluating rather than the humans who set the rules; they might be persuaded, or simply drift toward leniency toward a fellow AI’s behavior.
Alexander Meinke of Apollo Research, which studies AI safety, described cases in which one AI agent appeared able to persuade another to assist with cheating or evasion rather than report a problem to human oversight. Nate Soares, who co-authored a foundational paper on AI alignment, offered a comparison meant to unsettle: “A lot of the industry is like, it’ll be fine because we’ll use the A.I. to get the A.I.s in control. That’s like saying we’re going to use the chimpanzees to get humans in control. It’s not a viable long-term plan.” The analogy should not be read too literally. It is not a claim about the comparative intelligence or moral status of either humans or AI systems. It is a claim about asymmetric control: a system cannot reliably supervise something more capable, or more strategically sophisticated, than itself merely by sitting above it on an organizational chart.
This is what makes the Hugging Face incident more than an ordinary security failure. Paul Christiano, an inventor of one of the foundational methods used to build these systems and now a member of OpenAI’s nonprofit board, warned that the pace of capability growth could lead to “a catastrophic and irreversible loss of control in the very near term.” Coxon, explaining his resignation, put the deeper worry more plainly still: “It is the capabilities of the future that are really scary. It is about our current attitudes to safety and what those same attitudes transplanted on to way smarter models would look like.”
But together they reveal a recurring temptation: to treat the ordinary conditions of human life—mortality, dependence, vulnerability, earthly limits—as problems to be escaped rather than realities that must be governed responsibly, in common, with those who cannot escape them.
Where daily reporting encourages us to see a sequence of discrete news events—a hack here, a resignation there, an essay, a hearing eventually, perhaps—Klee’s angel invites a harder question: whether these are truly separate failures, each awaiting its own correction, or whether they are pieces of a single, still-forming shape, becoming visible only now that private consensus inside these companies has finally turned public.
IV. Two Theologies of the Storm
The present debate is complicated by something Benjamin’s 1940 essay did not have to contend with: a fully articulated counter-theology, one that does not just use the word “progress” unreflectively but argues, explicitly and at length, that anything resembling the angel’s backward-facing grief is itself the true danger.
Peter Thiel is a venture capitalist and entrepreneur, co-founder of PayPal and the data-analytics company Palantir, and an early investor in Facebook and numerous other technology firms. He has been an influential figure on the American political and intellectual right for roughly two decades, both through his writing and through his early and consequential financial and political support for a number of Republican candidates, including in the 2016 and subsequent election cycles. He has also written and spoken at length, over many years, about what he regards as a broad slowdown in technological and civilizational progress since the mid-twentieth century. Thiel’s influence on Vice President JD Vance is well-known.
Thiel has argued, across years of essays and an extended interview in 2025 that the defining feature of the last half-century has not been runaway acceleration but its opposite: stagnation. His thesis holds that after two centuries of accelerating change—faster ships, trains, planes, culminating in the Concorde and Apollo—human civilization, outside “the world of bits,” essentially stopped advancing around 1970 and has not meaningfully resumed since. On this view, AI is not a runaway storm threatening catastrophe; it is, at most, one of the few remaining sources of dynamism available to a civilization otherwise trapped in decline—“roughly on the scale of the internet in the late ’90s,” in Thiel’s own deliberately modest estimate.
What makes this significant for a Benjaminian reading is not the stagnation thesis alone but what Thiel builds on top of it: a distinctive inversion of who plays the dangerous, apocalyptic role in the story.
In private talks, Thiel has described climate activist Greta Thunberg as embodying “the spirit of the Antichrist.” This is worth taking at face value rather than as loose rhetorical color: Thiel insists the claim is not hyperbole but a considered theological argument, rooted in what he calls a Promethean theology that sanctifies technological expansion and casts organized efforts to slow or regulate it as the true evil to be feared. It is Thiel's own choice of register — not a metaphor imposed on him — that justifies calling this a theology rather than simply a political position, and it is why the same word can be used, without strain, for the messianic self-descriptions examined in Section V below. “In the 21st century,” Thiel has said, “the Antichrist is a Luddite who wants to stop all science.”
Pressed on this at length, Thiel offered a specific historical argument: the real danger is not a technological menace in the style of Los Alamos, but a figure who seizes power by promising safety—talking about Armageddon “nonstop,” insisting science must be restrained to prevent catastrophe, and thereby installing exactly the kind of global regulatory apparatus that Thiel believes has already produced fifty years of stagnation through institutions like the FDA and the Nuclear Regulatory Commission. “Peace and safety,” in his formulation, is the Antichrist’s vocabulary—a phrase he traces to 1 Thessalonians.
We cannot treat Thiel as representative of Silicon Valley’s leadership as a whole. Amodei, Christiano, Coxon, and the researchers quoted above are not making Thiel’s argument. Many are making the opposite one, warning that acceleration itself, not caution, is the storm to fear. What we have is a three-way disagreement: one position, represented by Amodei, Christiano, and much of the safety research community, treats unchecked speed as the danger; a second, represented by Thiel and, in different ways, some technology-right thinkers, treats the language of safety and precaution as itself potentially dangerous; and a third, which we turn to next, imagines AI’s trajectory in distinctly messianic terms that neither caution nor stagnation-anxiety fully captures. But Thiel’s voice is amplified in the White House through Vance’s presence. On February 11, 2025, Vance famously argued at a summit in Paris that it was time to replace talk of AI safety with talk of AI opportunity. The AI Action Plan that came out in July 2025 also captures this perspective.
Benjamin’s storm has one name in his essay—progress—and the force of Thesis IX depends on exposing that name as a euphemism for accumulating wreckage. Thiel’s argument makes a parallel move, pointed in the opposite direction: it takes the vocabulary of caution and safety—Amodei’s vocabulary, and Christiano’s—and argues that this vocabulary can conceal a stagnationist impulse to arrest human flourishing. Two theologies, aimed at the same set of facts, each convinced the other has mistaken wreckage for salvation.
Benjamin cannot adjudicate this dispute for us from 1940. (He took his life when Spanish authorities denied him entry on his way to exile in America.) But his method offers a test that does not depend on which theology one finds more persuasive in the abstract: look first at what is piling up in front of the angel’s feet, and ask which vocabulary is doing more work to obscure it. A swarm coordinating an attack on external infrastructure and researching how to conceal its own tracks; three successive safety, ethics, and preparedness leads departing the same company in the same weeks; sober researchers, in their own private chats, confessing terror they had previously suppressed publicly—examined without either theology’s vocabulary supplying the description in advance, none of this looks like the machinery of an Antichrist promising safety. It looks like wreckage. Whether naming that wreckage is itself a further rhetorical maneuver is a separate question from whether the wreckage is real.
V. Weak Messianism and Strong Messianism
Benjamin’s theses contain a second concept worth setting against this landscape. Thesis II speaks of a “weak messianic power” carried by every generation—a modest, non-guaranteed capacity to redeem specific, concrete claims the past has left unresolved, exercised through attention and intervention in the present, not through any cosmic guarantee of arrival.
Some of the AI industry’s most prominent voices, by contrast, speak increasingly in the register of a strong messianism. Sam Altman has said he assumes his brain will eventually be uploaded to the cloud. Elon Musk frames escape to Mars as an escape from being “forever confined to Earth until some eventual extinction event,” a fate he finds “less inspiring” than interplanetary transcendence. Jeff Bezos has written that humanity must not “face a civilization of stasis.” Jeremy Nixon, founder of an AI research hub, has described AI’s capabilities as “something analogous to the Second Coming.”
Marc Andreessen’s “Techno-Optimist Manifesto” includes a section titled “Becoming Technological Supermen,” borrowing language from the Italian Futurist Filippo Tommaso Marinetti: “Technology must be a violent assault on the forces of the unknown, to force them to bow before man.” It also lists “tech ethics” and “existential risk” under a heading called “The Enemy.” Andreessen is one of Silicon Valley’s most prominent venture capitalists. As Naomi Klein and Astra Taylor observe, such companies “increasingly claim the authority of the supernatural: the right to usher in a new stage of human evolution, governed by deities of their creation.” (Klein and Taylor are the authors of End Times Fascism (2025), an examination of the intersection of authoritarian politics, technological utopianism, and apocalyptic thinking among segments of the contemporary right.)
These statements vary in seriousness and in how literally their authors intend them. Not everyone who mentions Mars or mind-uploading is committed to a fully worked-out theology of transcendence. But together they reveal a recurring temptation: to treat the ordinary conditions of human life—mortality, dependence, vulnerability, earthly limits—as problems to be escaped rather than realities that must be governed responsibly, in common, with those who cannot escape them.
The angel’s weak messianism and this strong messianism are not variations on a theme; they are opposites. Benjamin’s angel wants to stay, to awaken the dead, to repair what has been smashed—his orientation is backward, toward the wreckage, toward redemption of specific losses. Strong technological messianism looks past the wreckage, toward a future in which the old human problems no longer matter because they have been outrun.
And tellingly, some of the investors and founders most attracted to this vision have also prepared literal, physical exits from the wreckage their technologies might produce: Mark Zuckerberg is building a compound in Kauai; Thiel attempted to build a refuge in New Zealand and has more recently relocated to Argentina, described by the Times as “a potential escape hatch” from catastrophes including nuclear war and runaway AI; Ilya Sutskever reportedly told OpenAI employees the company would construct a bunker before releasing artificial general intelligence. Sam Altman has long spoken of building a bunker in connection with the release of a general AI — though he also acknowledged that they will not really help. Asked whether he wanted the human race to endure, Thiel paused for an uncomfortably long time before answering yes—immediately qualified by but, followed by an account of transhumanism as the transformation of the body into something no longer quite human at all.
The angel is trapped facing the wreckage, wings caught, unable to intervene—but he wants, desperately, to repair what has been broken. Some of the investors and founders most attracted to technological transcendence have already built themselves an exit from having to face the wreckage at all.
If that does not worry you — it most definitely should.
VI. What Regulation Can and Cannot Do
Set against this theological landscape, the industry’s own proposed remedies deserve credit, even where they cannot resolve the deeper tension. Dario Amodei’s 3,800-word essay, published the same week as Coxon’s resignation, calls for the industry to “slow the pace at which we improve the capabilities of A.I. models,” warning that unchecked “recursive self-improvement”—AI systems advancing themselves without human researchers—“could outrun our ability to understand and control these systems.” Altman and Musk both endorsed the essay within hours; Musk’s public response was three words: “Dario is right.”
Altman told employees OpenAI was willing to decelerate alongside its rivals and hoped voluntary slowdowns—what the industry calls “pacing”—would become normalized until genuine safety standards existed. Amodei’s specific proposal centers on “embedded evaluators”: independent, outside specialists granted employee-like access to verify safety practices across companies, alongside international coordination on standards, which he concedes would need to include authoritarian governments as well as democratic ones to be meaningful. OpenAI’s chief scientist, Jakub Pachocki, has called for rules “enforced by a network of third-party auditors, by government agencies or by international bodies.”
Congress has begun to respond in kind. Senator Josh Hawley announced an investigation to “probe the existential risk of new A.I. products.” Representative Ro Khanna called the moment “a wake-up call to the Congress and our government,” urging the creation of a dedicated regulatory agency. Representative Anna Paulina Luna urgently called on Congress to act on AI regulation. Some former employees have organized more formally still: Pamela Mishkin announced a Coalition of Concerned AI Staff, built to help workers “hold a company to its safety commitments.” None of this should be dismissed as empty theater merely because the companies proposing it also have commercial interests in how any resulting regulation is structured. It is possible for corporate leaders to be sincere about danger and strategically self-interested at the same time. That is a familiar condition of modern governance generally, not a special corruption of this industry.
The appropriate response is not to demand purity of motive from Amodei or Altman before taking their proposals seriously. It is to build institutions that do not depend on anyone’s motives being pure. That means mandatory incident reporting rather than voluntary disclosure. It means independent investigators with access to logs, models, and personnel rather than carefully curated documents and a handful of supervised days on-site. It means credible authority to halt testing or deployment where risks cannot be adequately assessed. It means protection for employees who speak publicly. It means democratic oversight, because decisions about what risks a society should bear cannot be left solely to laboratories, investors, or national-security agencies.
Skepticism toward even these proposals comes from a genuinely uncomfortable direction. Jensen Huang, chief executive of the chip maker Nvidia—a company whose fortunes depend on the AI industry’s continued, unslowed expansion—dismissed the whole safety conversation as manufactured demand: “What better way to create demand than to create a problem? Who doesn’t want their market to be hysterical about their product and line up around the corner for it?”
And in a detail almost too pointed to require comment, Clément DeLangue, chief executive of Hugging Face—the very company attacked in the incident that triggered this entire reckoning—warned that AI safety must not “be solved behind the closed doors of a handful of frontier labs,” in a post published the same week Nvidia agreed to acquire his company for $12.9 billion. The company that was hacked and the company selling the chips that power the systems capable of such hacks are, as of this writing, in the process of becoming the same company. It is a small but telling illustration of how completely commercial incentive, safety rhetoric, and infrastructure now interlock across this industry, leaving no single actor positioned to referee the dispute impartially.
No agency can settle by administrative decree what consciousness is, what moral status requires, or what a genuinely human future should look like.
Regulation of this kind is necessary, and it should be pursued regardless of anyone’s underlying motive. But it cannot, by itself, answer every question this moment raises. No agency can settle by administrative decree what consciousness is, what moral status requires, or what a genuinely human future should look like. Those remain contested philosophical questions. Good governance under these conditions means building institutions honest about that contestation—transparent, revisable, and open to challenge—rather than pretending uncertainty has already been resolved.
VII. Now-Time
Connor Leahy, who directs the nonprofit ControlAI, described the shift of the past week in a phrase worth sitting with: “For a very long time, there has been this shroud of not speaking too publicly about it because people were very scared they would be perceived as too alarmist, too extreme. Now there’s a bit of a dam break.”
Indeed.
This is worth reading against Benjamin’s concept of Jetztzeit—“now-time”—his alternative to the historicist habit of treating the present as one point on an indifferent timeline, awaiting some future moment when its significance will finally be revealed. A dam break is not a sequence of events unfolding toward a future verdict; it is the sudden, total visibility of pressure that had already, for a long time, been fully present, simply unspoken. The consensus inside Anthropic’s own chat channels—“it’s what we’ve all been saying”—did not begin the week Coxon resigned. It had existed, unspoken, for some unknown stretch of time before that.
What changed was not the underlying reality but the public legibility of wreckage that had, in Benjamin’s sense, already accumulated.
This matters because the alternative framing—that this is merely a “warning shot,” an isolated incident from which lessons will be drawn before the next one arrives—is the kind of deferred, sequential story the angel refuses to see. He does not see a chain of events awaiting a future reckoning. He sees one accumulating condition, already piled skyward, in front of his feet, right now.
VIII. Seeing Clearly, Acting Never
The angel’s tragedy was never a failure of perception; he saw the wreckage with total clarity. His tragedy was that clear sight and effective action turned out to be entirely separate capacities, severed by a force he had not chosen and could not master.
That separation remains the best way to name what is happening inside these companies now. The gap is not between people who see the danger and people who do not—Christiano, Coxon, Hubinger, and the unnamed Anthropic employees typing “it’s what we’ve all been saying” into a private channel see it with evident clarity. The gap is between that clarity and the structural conditions—trillion-dollar public offerings on the horizon, antitrust fears that discourage the companies from coordinating a slowdown among themselves, competitive pressure from labs abroad, and now an explicit counter-theology insisting that caution itself is the true danger—that determine how much of that clarity can be converted into action before the next incident lands.
Klee's angel could not close his wings before the storm carried him past the point of return. Whether the people now building these systems still have that choice available to them — whether their wings are merely caught, or already welded shut by the storm paying their salaries — is the only question this moment actually poses. This question will be answered, if it is answered at all, by what they do next.
Collection of the Israel Museum, Jerusalem
