The Clash Between the Pentagon and Anthropic: The Ethics of AI at Stake

In just four months, Anthropic—the AI company whose corporate identity relies on the promise of developing safe AI—has been at the center of one of the most illuminating conflicts regarding the future of this technology. This includes a legal battle against the U.S. government over the right to impose limits on its own creations (algorithmic systems) and the publication of a document admitting that those creations might soon start improving themselves.

These two narratives are intertwined. To grasp what Anthropic envisions for the future of AI, one must first understand the war it has waged against the Pentagon.

The first public indication of the conflict arose in January 2026 when specialized media reported that the Department of Defense and Anthropic clashed over the military application of its model, Claude. The tension traced back even further: the model had been used, unbeknownst to the company, in the U.S. military operation that resulted in the abduction of Venezuelan President Nicolás Maduro during a violent raid on Caracas.

“Anthropic reached out to Palantir (by Peter Thiel)—the intermediary that had facilitated the military’s access to Claude—to express its concerns about how the model had been used in the operation,” reported The Wall Street Journal and Axios in February 2026.

The company made it clear that its “red lines” were non-negotiable: an absolute prohibition on use in lethal autonomous weapons and domestic mass surveillance. The Pentagon responded with an ultimatum.

24 FEB

Defense Secretary Pete Hegseth set a deadline: Anthropic had to acquiesce by 5:01 p.m. on February 27 and accept unrestricted use of Claude for “all legal purposes,” including autonomous weapons and mass surveillance.

27 FEB

Anthropic rejected the ultimatum. President Trump ordered federal agencies to cease using the company’s products. Hegseth designated Anthropic a “supply chain risk”—a label previously reserved for companies tied to foreign adversaries like China.

1 MAR

OpenAI (CEO Sam Altman) signed an agreement with the Pentagon hours before the U.S. initiated war against Iran, without the restrictions that Anthropic had demanded. Even OpenAI’s CEO acknowledged that the timing “seemed opportunistic and careless.”

9 MAR

Anthropic sued the federal government in two different courts—California and the D.C. Circuit—arguing that it faced irreparable harm and a violation of the First Amendment.

26 MAR

Judge Rita F. Lin granted a preliminary injunction blocking the designation, concluding in her 43-page ruling that the government had taken “retaliatory actions” against Anthropic for making its position public—a classic form of unconstitutional retaliation.

8 APR

The D.C. Circuit Court of Appeals overturned the preliminary injunction, arguing that lifting it would “force the armed forces to extend their relationship with an undesirable supplier amid an active military conflict.”

1 MAY

The Pentagon announced agreements with seven rival companies—SpaceX, OpenAI, Google, Nvidia, Microsoft, Amazon Web Services, and Oracle—to deploy AI in classified networks. Anthropic is left out.

Negotiations at the Highest Levels of Power

The case has an outstanding dimension. Weeks after the most acute conflict, Anthropic CEO Dario Amodei was welcomed directly at the White House to discuss Mythos, a new model capable of identifying cybersecurity vulnerabilities. President Trump himself termed the meeting “productive” and stated that an agreement was “possible.”

The supply chain risk designation remains in effect. This illustrates the power that major tech companies wield and their close relationship with political power in a context marked by wars, economic crises, and climate challenges.

Nevertheless, what the conflict revealed is, according to some, a clash of irreconcilable visions: the Trump administration, committed to the unrestricted acceleration of AI, versus a company that has made ethical limits its core identity.

For Anthropic, capitulation would have meant destroying the foundation upon which its credibility rests. For others, less naive, it’s an all-out fight in the lucrative market of capitalist AI companies (the Big Tech) and for the cognitive control of humanity.

The Explosion of Artificial Intelligence: Humanity’s Existential Risk

Amid this context—banned by the Pentagon, suing the government, negotiating its return—Anthropic published this week (May 6-7) the research agenda for its new Anthropic Institute. The document contains an unprecedented admission in the history of the sector.

The company acknowledges having observed early signs that its AI systems are beginning to speed up the development of their future iterations, a phenomenon known as “recursive improvement.” And they put it in writing. In other words, AI systems could operate autonomously.

“By the end of 2028, it’s very likely that we will have an AI system to which you can say, ‘Make a better version of yourself,’ and it will simply do so completely autonomously” (Axios, May 7).

Jack Clark, co-founder of Anthropic and director of the institute, made this prediction in an interview from the company’s headquarters in San Francisco, estimating a greater than 60% chance that by the end of 2028, an AI model will have autonomously trained its successor.

For years, the term “intelligence explosion” belonged exclusively to philosophers and theorists of existential risk. This week, it appeared in an official document from Anthropic—one of the world’s most influential AI laboratories—and does not sound like an abstract warning anymore.

An intelligence explosion—in Clark’s definition—occurs when AI systems start improving at a breakneck pace. The consequences could be twofold: severe risks like cyber collapses or biological attacks, along with unprecedented opportunities in medicine, biology, and science.

The document questions how to expand the current channels of pharmaceutical development in light of the possible onslaught of drug candidates that these systems would generate.

The Nuclear Fire Drill

One of the most striking proposals in Anthropic’s document is to conduct “fire drills” in anticipation of a possible intelligence explosion: tabletop exercises that would test the response capabilities of lab executives, their boards, and governments. Labs don’t design drills for problems they believe are decades away.

The Philosophical Question

What Clark suggests is not merely a technical concern but one that is deeply philosophical and civilizational. Until now, advancements in artificial intelligence have required that individuals outside the system contribute ideas which are then incorporated into the development process. What changes with recursive improvement is that technology would generate those ideas internally.

“It has always been the case that humans outside of technology had to generate the ideas that they would then introduce into it. What if we have a technology that can generate its own ideas on how to improve itself? That’s a new concept,” explained Jack Clark, co-founder of Anthropic and director of the Anthropic Institute.

The document also proposes something unprecedented in the industry: the possibility of AI companies, in coordination with governments, operating as technological central banks, adjusting “dials” to regulate the speed of AI diffusion sector by sector. The analogy with monetary policy is no coincidence: it suggests that the spread of artificial intelligence may require deliberate macroeconomic management.

Additionally, Anthropic has committed to publishing monthly reports on how AI is reshaping work, described as “early warning signs” for significant changes in employment. They also pledged to publicly disclose when their tools start accelerating the development of their models.

Sincerity or Marketing Strategy?

Of course, there’s a sharper reading. Anthropic has built its corporate brand on the promise of being the responsible laboratory in a fast-paced sector. An institute linked to safety and transparency bolsters that positioning just as it approaches the launch of its next model.

Clark acknowledges this directly: “The motivation has always been to tell the complete story. I’m just trying to get ahead of what I believe to be the next big question and position Anthropic at the forefront of it.”

Yet, this strategic transparency does not diminish the underlying admission: a company at the forefront of technological development has officially stated that its systems show signs of becoming autonomous from their creators. And it has promised to reveal when that becomes clear. In a sector where public self-criticism is rare, that alone is a novelty.

More Philosophy…

However, citizens know that novice sorcerers do not stop, and states either observe passively or follow the path of amplification/projection of power (Machiavelli).

For to create an object that escapes your grasp and subsequently controls you is a process that Marx termed alienation—drawing on Hegel— in his “1848 Manuscripts”: you build an artifact, invest your work, your knowledge, your life into it. Initially, it obeys you. Yet one day, it slips from your control, becomes autonomous, starts to operate on its own. Then it is no longer you who controls it; it controls you. It claims your being, imposes its rhythm on you, dictates what you are worth, and what you were supposed to master becomes your master.

Thus, quietly, the creature emerges as an existential risk: it no longer threatens just your body, but your sense of being free. That is alienation in history. The circulating capital—expropriated, exploited labor accumulated as money—is the same. And the digital Frankenstein, its outcome. Marx articulated the concepts before monitors existed.

Leopoldo Lavín Mujica
Based on articles from Axios, WSJ, The New York Times.

La entrada The Clash Between the Pentagon and Anthropic: The Ethics of AI at Stake se publicó primero en El Ciudadano.