The Company That Blinked

Anthropic’s Safety Capitulation and the Price of Survival

Eight days ago, this publication called Anthropic “the only company standing.” The headline felt earned. Three AI firms had capitulated to Pentagon pressure. One had not. Dario Amodei had drawn two lines in the sand: no mass surveillance of American citizens, no AI-controlled autonomous weapons. His resistance was being celebrated by constitutional scholars, AI researchers, and everyone who understood that what the Pentagon was demanding was not a contractual adjustment but a restructuring of institutional accountability under the cover of national security.

On February 24, 2026, Anthropic published a blog post titled “Responsible Scaling Policy v3.”

Read it carefully, because the sentence that matters most is the one that disappears.

The founding commitment the one that made Anthropic structurally different from every other AI company racing toward deployment was this: if our models’ capabilities outstrip our ability to ensure their safety, we will pause. Not slow down. Not reassess. Pause. That commitment is gone. In its place, Anthropic offers “public goals” that the company will “openly grade our progress towards.” On the same day, Defense Secretary Pete Hegseth sat across from Amodei and gave him a 5:01pm Friday deadline: remove AI safeguards or lose a Pentagon contract worth up to $200 million. He also threatened to invoke the Defense Production Act and designate Anthropic a supply chain risk, a classification typically reserved for foreign adversaries.

The timing is not coincidental. Anyone who says otherwise is not reading the sequence.

What Changed Between February 17 and February 24

When this publication first covered the Anthropic-Pentagon conflict, the company was resisting. The Venezuela operation had exposed the mechanics of coercion: the Pentagon used Claude in a military operation, then demanded that Anthropic agree in advance that any use the government deemed “lawful” was acceptable. Tensions escalated. The contract was threatened. The company held its ground.

What has shifted is not the coercion. The coercion is still there, explicit and intensifying. What shifted is the institutional structure Anthropic built to resist it.

The original Responsible Scaling Policy was not an aspiration. It was a pre-commitment device. The company’s founders understood something about institutional behavior under pressure: good intentions are insufficient when the financial stakes are existential. A company facing contract termination, government blacklisting, and competitive disadvantage cannot rely on its values alone to hold the line. The binding commitment was designed to remove the option to rationalize. If safety cannot keep pace, you stop. The decision is made in advance, when the pressure is not yet present.

The Stated Rationale and What It Obscures

Chief Science Officer Jared Kaplan’s explanation for the policy change is factually accurate and structurally incomplete.

He is right that Anthropic’s original “race to the top” strategy failed. The company hoped that by establishing rigorous safety standards and making them public, other AI firms would face reputational and commercial pressure to follow. They did not. OpenAI, Google, and xAI all agreed to Pentagon terms allowing their models to be used for “all lawful purposes” in military contexts, without the restrictions Anthropic maintained. The industry did not follow Anthropic’s lead. It watched Anthropic maintain costly commitments alone and calculated that the competitive disadvantage was not worth the principle.

There is an additional complication that honest reporting requires naming: internal discussions about revising the RSP began in February 2025, nearly a year before the Pentagon ultimatum. The policy was not drafted overnight in response to Hegseth. It was a long internal process, and Anthropic’s own researchers describe it as a genuine reckoning with what the original framework could and could not achieve.

This matters. It means the RSP v3 is not purely a capitulation to external pressure. It is also a genuine reassessment of whether a unilateral pause commitment, in a market where three competitors refused to adopt comparable constraints, could function as intended.

And yet: a company publishing a weakened safety framework on the same day its CEO receives a Pentagon ultimatum has an obligation to name that connection directly in its public communications. Anthropic did not. The blog post frames the change as a considered evolution of policy in response to industry dynamics. It does not mention Hegseth. It does not mention Venezuela. It does not mention the Friday deadline. This silence is not neutral. It is a choice about what the public record will say.

What the Venezuela Dispute Actually Documents

The dispute that triggered this crisis requires precise reconstruction, because both sides have conflicting accounts.

The documented facts: Claude was used in the January operation that captured Venezuelan President Nicolás Maduro, through Anthropic’s partnership with Palantir. This was confirmed by multiple sources to Axios. After the operation became known, a rupture emerged. According to a senior Pentagon official who spoke to NBC News, “a senior executive from Anthropic communicated with a senior Palantir executive, inquiring as to whether their software was used for the Maduro raid,” and the Palantir executive “was alarmed that the question was raised in such a way to imply that Anthropic might disapprove of their software being used during that raid.”

Anthropic flatly denied this account. A company spokesperson told NBC News that Anthropic “has not discussed the use of Claude for specific operations with the Department of War” and has not discussed such matters with industry partners “outside of routine discussions on strictly technical matters.” At the Hegseth meeting on February 25, Amodei personally denied that Anthropic raised concerns about the Maduro raid with Palantir, telling Hegseth the company’s red lines have never prevented the Pentagon from doing its work or posed an issue for anyone operating in the field.

Two contradictory accounts. One from anonymous Pentagon officials. One from the company’s CEO directly. The Pentagon account became the stated pretext for the ultimatum. Anthropic’s denial has not been publicly disputed with additional documentation.

What is not disputed: Claude is currently the only AI model operating inside the Pentagon’s classified systems. OpenAI, Google, and xAI have agreed to “all lawful purposes” terms for unclassified systems. Anthropic was the holdout. The conflict predates the Maduro dispute by months: a senior administration official told Axios that defense officials “have been frustrated with Anthropic for some time, and embraced the opportunity to pick a public fight.”

What Mrinank Sharma Told Us

On February 9, Mrinank Sharma resigned as head of Anthropic’s Safeguards Research Team. His departure, announced in a letter posted publicly on X and viewed more than ten million times, described his two years at the company in detail: work on understanding AI sycophancy, developing defenses against AI-assisted bioterrorism, writing one of the first AI safety cases.

Then he wrote: “Throughout my time here, I’ve repeatedly seen how hard it is to truly let our values govern our actions. I’ve seen this within myself, within the organization, where we constantly face pressures to set aside what matters most, and throughout broader society too.”

Sharma’s letter is unusually philosophical and does not make specific accusations. Multiple outlets noted its cryptic, literary character. He cited poets. He announced plans to pursue a poetry degree. He does not accuse Anthropic of specific wrongdoing in the Venezuela dispute or the Pentagon negotiations.

What his letter documents is this: the person operationally responsible for Anthropic’s safety research saw, from inside the institution, a persistent pattern of pressure to set aside values when they conflicted with other demands. He chose to leave rather than continue navigating that pattern.

His resignation came fifteen days before the RSP v3 was published. The sequence should be read as context, not as direct causation.

The Two Red Lines: How Long Do They Hold?

To its credit, Anthropic is holding two positions. No AI-controlled autonomous weapons operating without human oversight. No mass domestic surveillance of American citizens. A source familiar with the Hegseth meeting confirmed to CNN that Amodei held these positions directly. AI researchers and civil liberties organizations celebrated this publicly. The Brennan Center’s Amos Toh wrote that the law is “not keeping up with how quickly the technology is evolving” but that the Pentagon “doesn’t have a blank check.”

The celebration is not misplaced. These are real commitments with real stakes.

But the structural question is different from the motivational one. The binding commitment to pause development existed because good intentions are insufficient when the financial stakes are high enough. Anthropic’s own published research documents how their models could be used for blackmail under certain conditions. The company’s founders built a pre-commitment device because they understood the danger of making these decisions in real-time under financial pressure.

The pre-commitment device is gone. The two red lines remain, maintained now by resolve rather than structure.

Pentagon officials have already told CNN that they consider the line between lawful and mass surveillance too “murky” to enforce, and that the distinction between human-controlled and autonomous targeting is too “ambiguous” as a boundary. The government’s position is that if the executive branch deems an operation lawful, any contractor inserting restrictions on that operation is creating an operational liability.

This is the mechanical operation of institutional erosion. Remove the structural protection first. Then argue the remaining commitments are unworkable. Then apply the next round of pressure.

The Financial Architecture

The Pentagon contract worth “up to $200 million” is the visible number in a much larger architecture.

Amazon has invested $8 billion in Anthropic and serves as the company’s primary cloud and training partner through AWS. Google has invested approximately $3 billion total, with additional cloud infrastructure commitments worth tens of billions through 2029. The February 12 Series G funding round closed at $380 billion valuation, up from $183 billion in September 2025, with participants including Singapore’s sovereign wealth fund GIC, Coatue, Founders Fund, Microsoft, Nvidia, and the Qatar Investment Authority. Annualized revenue has reached $14 billion, growing more than 10x annually for three consecutive years.

Eight of the ten largest companies in the United States use Claude. Claude Code alone has $2.5 billion in annualized revenue. The $200 million Pentagon contract is, as Axios noted, “a small fraction of Anthropic’s $14 billion in annual revenue.”

This financial picture reframes the leverage question. The Pentagon is not threatening a company that depends on government contracts. It is threatening a company whose entire valuation rests on a single claim: that Anthropic’s safety-first positioning makes Claude uniquely trustworthy for high-stakes enterprise use. The supply chain risk designation, which would require any company with Pentagon contracts to certify it does not use Anthropic’s products, is the genuine weapon. Not the contract termination.

If the supply chain risk designation spreads to federal vendors and contractors, the exposure touches a significant portion of the enterprise clients that drive Anthropic’s revenue. Government contracts law expert Franklin Turner, of McCarter and English, told Reuters the scenario is “unprecedented and will almost certainly trigger a raft of downstream litigation.” The Defense Production Act invocation would face immediate legal challenge on whether AI model behavior constitutes an “industrial resource” under a statute designed for manufacturing and supply chains.

The Pentagon is betting on the threat. The legal basis for the most aggressive options is contested.

The $20 Million That Names the Stakes

On the same day Anthropic closed its $30 billion Series G funding round, February 12, the company announced a $20 million donation to Public First Action, a bipartisan 501(c)4 established to back congressional candidates who support AI regulation.

The two announcements appearing on the same day is not incidental to the institutional story. Anthropic is simultaneously raising money at historic valuations and spending it on political candidates who would impose legal constraints on AI development, including constraints on Anthropic itself. The company explicitly said this in the press release: “effective AI governance means more scrutiny of companies like ours, not less.”

David Sacks, Trump’s AI and crypto czar, responded by calling this a “sophisticated regulatory capture strategy based on fear-mongering.” The White House signed an executive order establishing a single federal AI framework that preempted state-level regulation, directly countering what Public First Action is organized to support.

Within two weeks of these announcements, Hegseth was in a room with Amodei delivering a deadline.

Why This Matters to Everyone Reading from the Global South

This conflict is being covered as an American corporate story. It is not.

The same AI infrastructure being integrated into Pentagon classified systems is being deployed across Global South military partnerships, intelligence cooperation agreements, and security assistance programs. The pattern of military AI deployment follows the pattern of American military influence: it goes where American strategic interests go.

When the Pentagon establishes that private AI companies can be disciplined into removing safety constraints through contract pressure and supply chain risk designations, it establishes a template. That template travels. The intelligence architectures built for the Venezuela operation will be adapted for the next operation, in the next country, through the next bilateral security agreement.

Pakistan understands what “all lawful purposes” means when an executive branch defines lawfulness unilaterally. The drone strike program ran for years on precisely this logic. The legal authorization was classified. Casualties were designated enemy combatants after the fact. The algorithm selecting targets was proprietary. Accountability was nonexistent.

What Anthropic’s original safety framework represented was an attempt by a private institution to insert a checkpoint into that pipeline. Not a refusal to serve national security, but a refusal to agree in advance that all executive branch decisions are lawful and all applications of its technology are therefore acceptable. The right to ask questions. The right to know how Claude was used.

That checkpoint is structurally weakened. The two red lines Anthropic is holding are the ones most visible to domestic American audiences: surveillance of American citizens, autonomous weapons. The applications that matter most to the Global South, including intelligence synthesis for regime change operations, were never among the explicitly stated restrictions.

What Comes Next

Anthropic has not fully surrendered. Amodei held the two lines in a direct confrontation with the Defense Secretary. Legal experts say the Pentagon’s most aggressive options face immediate judicial challenge. The company has $14 billion in annualized revenue, $380 billion in enterprise valuation, and clients whose procurement officers care deeply about whether Claude’s safety commitments are credible.

The pressure is not over. The Friday deadline passed as this article was being written. Whether Hegseth follows through, whether the Defense Production Act is actually invoked, whether the supply chain risk designation is issued, these remain open questions whose answers will shape the next phase of this story.

What is already documented is this: the pre-commitment device is gone. The “binding ourselves to the mast” architecture, as Anthropic’s own researcher described the original RSP, has been replaced with public goals and self-grading. The institution that built the strongest safety framework in the industry concluded, after nearly a year of internal deliberation and under simultaneous external pressure, that the framework as designed could not be sustained.

Amodei wrote in “The Adolescence of Technology” that the central civilizational question is whether we can develop wisdom in proportion to power. His company is being forced to answer that question not in essays but in contract negotiations, blog posts, and meetings with defense secretaries who call its principles “woke AI.”

The company that said no in January is negotiating what no means in February.

The structural protection that made that no durable is gone.

What remains is the character of the people holding the line, and their ability to hold it without the architecture that was designed to make the decision for them.

History does not have a good record on that bet.