Anthropic Said No. The Pentagon Blacklisted Them. Then OpenAI Got the Exact Same Deal.

On Friday, February 27, 2026, at 5:01 PM Eastern Time, a deadline expired.

President Trump posted on Truth Social ordering every federal agency to stop using Anthropic's products. Defense Secretary Pete Hegseth designated Anthropic — the company that built Claude, the first frontier AI model deployed on classified military networks — a "Supply-Chain Risk to National Security." Every defense contractor, supplier, and partner that does business with the United States military was barred from conducting commercial activity with Anthropic.

The offense: Anthropic refused to remove two safety restrictions from Claude. No mass surveillance of American citizens. No fully autonomous weapons without a human in the loop.

Hours later, OpenAI CEO Sam Altman announced that his company had signed a Pentagon deal for classified network access. The terms included prohibitions on mass surveillance and autonomous weapons — the exact protections Anthropic was punished for demanding.

This is the complete story of how the most consequential AI policy confrontation in history unfolded over five weeks, what it reveals about the relationship between artificial intelligence and state power, and why the outcome may be worse than either side's supporters understand.

The Contract That Started Everything

In July 2025, Anthropic signed a contract valued at up to $200 million with the Pentagon. Through its partnership with Palantir Technologies, Claude became the first frontier AI model deployed on classified U.S. government networks — supporting intelligence analysis, threat modeling, cyber operations, and defense planning.

The contract was considered a milestone for both Anthropic and the broader AI industry. A company founded on the principle that AI development should prioritize safety above all other considerations had chosen to engage with the most consequential customer in the world — the United States Department of Defense.

Anthropic's logic was straightforward: if AI would inevitably be used by militaries, it was better for a safety-focused company to provide that technology with guardrails than to cede the market to competitors with fewer restrictions. The contract included specific limitations on Claude's use — boundaries that Anthropic considered non-negotiable.

Those boundaries would become the fulcrum of a crisis that threatened to reshape the entire relationship between Silicon Valley and the national security state.

The Maduro Catalyst

The crisis began with a raid in Caracas.

In January 2026, U.S. Special Operations Forces captured Venezuelan President Nicolás Maduro in an operation that reportedly used Claude during active military operations. The details that emerged were stark: 83 people were killed during the raid, including 47 Venezuelan soldiers. Claude was deployed through the Palantir partnership — not for planning alone, but reportedly during the operation itself.

An Anthropic executive reportedly reached out to a Palantir executive to ask whether Claude had been used in the Maduro operation — raising concerns about their AI being deployed in an operation involving lethal force. Palantir reported the inquiry to the Pentagon. The Pentagon interpreted it as evidence of unreliability.

The question was whether an AI company had the right — or the obligation — to ask how its technology was being used in combat. Anthropic believed it did. The Pentagon believed the question itself was disqualifying.

Hegseth's "AI-First Warfighting Force"

On January 9, 2026, Defense Secretary Pete Hegseth had already laid the groundwork for confrontation. In a memorandum to senior Pentagon officers, he announced a push toward an "AI-first warfighting force" and ordered that all Department of Defense AI contracts incorporate standard "any lawful use" language within 180 days.

The directive was explicit: AI models should be used "for all legitimate military purposes, free from usage policy constraints" set by individual AI companies.

The "any lawful use" language sounds reasonable in isolation. Its implications become visible only when you examine what is and is not lawful under current U.S. policy.

DOD Directive 3000.09, updated in January 2023, governs autonomy in weapons systems. It requires "appropriate levels of human judgment over the use of force" but does not define what "appropriate" means. It does not establish minimum thresholds for human involvement. It does not prohibit fully autonomous lethal engagement. Autonomous weapons are not illegal under U.S. law. The United States voted against a UN General Assembly resolution calling for a legally binding agreement on lethal autonomous weapons systems.

"All lawful purposes" includes autonomous weapons. It includes mass surveillance. That is specifically why Anthropic drew the line.

The Confrontation Week

Tuesday, February 24: The Face-to-Face Ultimatum

Hegseth met Dario Amodei in person and delivered the terms. Allow Claude to be used "for all lawful purposes" with no company-imposed restrictions. The deadline: 5:01 PM ET, Friday, February 27.

The consequences of refusal were unprecedented for an American technology company:

1. Designation as a "Supply-Chain Risk to National Security" — a classification normally reserved for companies connected to foreign adversaries like China's Huawei
2. Invocation of the Defense Production Act — a Korean War-era statute designed for steel mills and tank factories — to force compliance

Amodei noted the threats were "inherently contradictory: one labels us a security risk; the other labels Claude as essential to national security."

On the same day, Anthropic released version 3.0 of its Responsible Scaling Policy. The timing would become deeply controversial.

The xAI Deal: February 23

One day before the Anthropic ultimatum, the Pentagon signed Elon Musk's xAI to deploy Grok in classified military systems. xAI agreed to the "all lawful purposes" standard without restriction — the exact terms Anthropic refused. Grok was given access to systems used for "some of the military's most sensitive intelligence analysis, weapons development and battlefield operations."

The strategic logic was transparent. By signing xAI first, the Pentagon demonstrated that compliant alternatives existed, reduced Anthropic's leverage, and established the unrestricted standard as an industry norm.

Wednesday, February 25: The Blacklist Begins

The Pentagon took concrete preparatory steps. Officials asked defense contractors Boeing and Lockheed Martin to assess their "exposure" to Anthropic products — a clear signal that the supply-chain risk designation was being prepared regardless of whether Anthropic complied.

Hegseth publicly framed Anthropic's safety restrictions as "ideological constraints," recasting technical safety commitments as culture war artifacts. The "woke AI" label served a specific political function: creating public permission for aggressive government action against safety commitments by associating those commitments with an ideology the administration's base already opposed.

Wednesday-Thursday: The "Final Offer"

Pentagon officials sent what they called a "best and final offer." Anthropic rejected it, stating the new contract language "made virtually no progress on preventing Claude's use for mass surveillance of Americans or in fully autonomous weapons."

A critical moment came when the Pentagon's technology chief posed a hypothetical: if a nuclear-armed ICBM were hurtling toward the United States with only 90 seconds to respond, and Claude were the only way to trigger a missile defense system, would Anthropic's guardrails block it?

Anthropic offered to allow Claude's use for missile defense — but the Pentagon wanted broader language with no restrictions whatsoever. The nuclear scenario was a negotiating tactic, not a genuine operational concern.

Thursday, February 26: Amodei Goes Public

Dario Amodei published a formal statement that read like a man who had already accepted the consequences:

"Anthropic understands that the Department of War, not private companies, makes military decisions. In a narrow set of cases, we believe AI can undermine, rather than defend, democratic values. Some uses are also simply outside the bounds of what today's technology can safely and reliably do."

"These threats do not change our position: we cannot in good conscience accede to their request."

He noted that Anthropic was "the first frontier AI company" to deploy in classified networks and had forfeited "several hundred million dollars" in revenue by cutting off Claude access to entities linked to the Chinese Communist Party.

The Pentagon's response was personal. Emil Michael, Under Secretary of Defense for Research and Engineering and a former Uber executive, called Amodei a "liar" with a "God complex" who was "ok putting our nation's safety at risk."

Friday, February 27: The Hammer Falls

At 5:01 PM, the deadline expired. The consequences were immediate and comprehensive:

Trump posted on Truth Social ordering every federal agency to immediately stop using all Anthropic products, with a six-month phase-out period. Hegseth designated Anthropic a "Supply-Chain Risk to National Security." The Pentagon severed the $200 million contract. And Hegseth posted on X: "Effective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic."

The OpenAI Paradox

Then, within hours, the story inverted.

OpenAI CEO Sam Altman announced that his company had signed a deal with the Pentagon for deploying models in classified networks. The critical detail was in the terms:

"Two of our most important safety principles are prohibitions on domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems. The DoW agrees with these principles, reflects them in law and policy, and we put them into our agreement."

The Pentagon agreed to OpenAI's red lines. The same red lines it had refused to accept from Anthropic. The same protections it had blacklisted a company for demanding.

Altman also stated publicly that he did not "personally think the Pentagon should be threatening DPA against these companies."

The implication was devastating for the Pentagon's position. Either the safety restrictions were genuinely unacceptable for national security — in which case the Pentagon should not have agreed to them with OpenAI — or they were always acceptable, and Anthropic was punished not for its policy positions but for its refusal to submit without negotiation.

The most plausible interpretation is the most uncomfortable one: the blacklisting was retaliatory. Anthropic was punished for questioning the Maduro raid, for insisting on formal contractual protections rather than informal assurances, and for making the confrontation public. The safety restrictions were the stated reason. They were not the actual reason.

The Responsible Scaling Policy: Separate Crisis, Same Week

In a development with deeply unfortunate timing, Anthropic released version 3.0 of its Responsible Scaling Policy on February 24 — the same day Hegseth delivered his ultimatum.

What Was Removed

The most significant change was the elimination of Anthropic's founding safety pledge: the commitment to never train an AI system unless it could guarantee in advance that safety measures were adequate. This "training pause" commitment had been the company's defining policy differentiator — the concrete mechanism that distinguished Anthropic's approach from every other AI lab.

Under RSP v3.0, Anthropic will only "delay" — not halt — development if it believes it has a "significant lead" on competitors. The binding promise to stop if capabilities outstripped safety measures was replaced with a promise to be "transparent" about whether the company was meeting its safety goals.

Anthropic's Justification

Chief Science Officer Jared Kaplan explained the reasoning: "We felt that it wouldn't actually help anyone for us to stop training AI models. We didn't really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments... if competitors are blazing ahead."

The company cited three forces driving the change:

1. A "zone of ambiguity" around capability thresholds that made bright-line commitments impractical
2. An "increasingly anti-regulatory political climate" that reduced the value of voluntary restraint
3. Requirements at higher capability levels that demanded "industry-wide coordination" rather than unilateral action

The Timing Problem

Anthropic insists the RSP revision was planned independently of the Pentagon confrontation. This may be true. It does not matter. The perception — reinforced by extensive media coverage — is that Anthropic simultaneously held a hard line against the Pentagon on specific use restrictions while softening its most fundamental safety commitment on model development.

The result is a company that appears to have defended one guardrail while dismantling another, leaving observers uncertain about which principles are genuinely non-negotiable and which are subject to competitive pressure.

For detailed analysis of the RSP changes and their implications, see The Great Unalignment.

The Industry Responds

"We Will Not Be Divided"

On the day of the blacklisting, employees at Google and OpenAI published an open letter titled "We Will Not Be Divided." More than 300 Google employees and 60 OpenAI employees signed.

The letter accused the Pentagon of trying to "coerce Anthropic" and attempting to "divide each company with fear that the other will give in." Signatories urged their employers to "put aside their differences and stand together" to uphold the boundaries Anthropic asserted.

Google DeepMind Chief Scientist Jeff Dean wrote publicly: "Mass surveillance violates the Fourth Amendment and has a chilling effect on freedom of expression."

The letter represented something rare in the AI industry: competing companies' employees publicly supporting a rival on a matter of principle. The competitive dynamics that normally prevent solidarity — fear that one company's restraint becomes another's advantage — were temporarily suspended by the shared recognition that the precedent being set would apply to all of them.

Congressional Response

The congressional response was notable for its bipartisan character:

• Sen. Thom Tillis (R-NC) called the public dispute "sophomoric," saying consequential decisions about AI in warfare should not be made through ultimatums and social media posts
• Sen. Mark Kelly (D-AZ): "DOD is trying to strong-arm Anthropic into providing every tool they have to surveil U.S. citizens. That is unconstitutional. That's not the role of the Department of Defense"
• Rep. Ro Khanna (D-CA): "Good for Anthropic." He asked: "Would you want them to have tools of AI to figure out whether you have a gun, to figure out what church you're going to?"

Congressional members from both parties called for the Friday ultimatum to be stayed and for Congress to work with the administration and industry to develop "clear rules of the road" for military AI use.

No concrete legislation has been introduced as of February 28. Bipartisan concern has not yet translated into bipartisan action.

The Legal Battle Ahead

Can the Pentagon Do This?

Legal experts are broadly skeptical of the government's position. The Defense Production Act is a Korean War-era statute designed for compelling steel mills and munitions factories to prioritize government orders. Mapping it onto an AI software dispute is legally unprecedented.

Lawfare published detailed analysis noting that "different demands raise very different legal questions" and that experts "don't expect the government to prevail" because "it seems very out of bounds under the text of the law."

In the end, the DPA was not invoked. The administration chose the supply-chain risk designation and government-wide ban instead — a potentially more effective punishment that avoids the legal vulnerabilities of compelling a company to modify its software.

Anthropic's Challenge

Anthropic has announced it will challenge the supply-chain risk designation in court, arguing:

• The designation is "legally unsound" and sets "a dangerous precedent for any American company that negotiates with the government"
• The Secretary "does not have the statutory authority" for the broad ban Hegseth announced
• Under federal law, the supply-chain risk designation would only apply to Claude's use in Department of Defense contracts, not how defense contractors use Claude for other customers

The legal challenge could take months or years to resolve. In the interim, the practical effect of the designation — defense contractors avoiding Anthropic to protect their own Pentagon relationships — will reshape the market regardless of the eventual court outcome.

For our prediction on whether the Defense Production Act will ultimately be invoked against an AI company, see U.S. Defense Production Act vs. AI Companies.

What Actually Changed — And What Didn't

Anthropic's Military Red Lines: Held

Anthropic did not remove its guardrails on autonomous weapons or mass surveillance. The company accepted the consequences — loss of the $200 million contract, supply-chain risk designation, government-wide ban, potential loss of defense contractor relationships — rather than comply.

This makes Anthropic the first major technology company in American history to be designated a national security supply-chain risk for maintaining safety restrictions on its own product. Whether this represents principled courage or strategic miscalculation depends entirely on whether you believe the restrictions matter.

The RSP Training Pause: Removed

Separately, Anthropic did remove its founding commitment to halt AI development if capabilities outstripped safety measures. This change was framed as pragmatic — competitors advancing without restraint made unilateral pauses ineffective — but it represents a real erosion of the most concrete safety mechanism in the industry.

The net result is a company that held its most visible guardrails (military use restrictions) while quietly dismantling its most fundamental one (the training pause). Both positions have their own logic. Together, they create an uncomfortable asymmetry.

The Pentagon's Position: Contradicted By Its Own Actions

The OpenAI deal is the most damaging fact for the Pentagon's narrative. If the safety restrictions Anthropic demanded were genuinely incompatible with national security, the Pentagon should not have accepted identical restrictions from OpenAI hours later. The fact that it did reveals the confrontation was never primarily about the restrictions themselves.

The Pentagon now has three AI relationships with three different standards:

• xAI/Grok: No restrictions. Full "all lawful purposes" access to classified systems
• OpenAI/GPT: Restrictions on mass surveillance and autonomous weapons (identical to what Anthropic demanded)
• Anthropic/Claude: Blacklisted for demanding restrictions that OpenAI received

This is not a coherent policy. It is a punishment.

The Bigger Picture

The Precedent Problem

The most consequential aspect of this confrontation is not what happened to Anthropic. It is what happens next.

Every AI company now knows the cost of maintaining safety restrictions against government demands: loss of government contracts, supply-chain risk designation, potential DPA invocation, and a public campaign framing safety commitments as "woke" ideology.

Every AI company also knows the cost of compliance: nothing. xAI accepted unrestricted access and was immediately rewarded. OpenAI accepted restrictions but framed them cooperatively rather than confrontationally and was rewarded.

The lesson is not about whether to have guardrails. It is about whether to fight publicly for them. Anthropic's mistake, in the realpolitik reading, was not its position but its posture — making the confrontation visible rather than negotiating quietly to the same endpoint OpenAI achieved.

The Defense Industrial Base Fallout

The supply-chain risk designation creates cascading effects throughout the defense industrial base that extend far beyond Anthropic's direct revenue.

Defense contractors operate in a compliance-driven environment where even the appearance of association with a designated entity creates procurement risk. Companies that integrated Claude into intelligence analysis workflows, cybersecurity tools, or logistics optimization systems now face forced migration timelines — replacing a model that was specifically selected for its capabilities and deployed on classified networks.

The six-month phase-out period gives these organizations time to transition, but transitions of classified AI systems are not simple license swaps. Models are fine-tuned on specific datasets, integrated into custom pipelines, and validated against security standards that took months to achieve. Switching to GPT-4 or Grok requires re-validation, re-certification, and re-deployment across every system that currently uses Claude.

The irony is that Grok — the replacement the Pentagon is promoting — has no track record in classified environments. Claude was chosen precisely because it passed the security evaluations other models had not. The government is now replacing a vetted, proven system with an untested alternative because the vetted system's creator insisted on maintaining two specific restrictions that another company's identical restrictions were deemed acceptable.

Several defense contractors have privately expressed frustration with the designation, according to reporting by Bloomberg and Semafor. They chose Anthropic for technical capabilities and safety characteristics that are directly relevant to classified work — and are now being punished for that choice by the same customer that originally approved it.

The Surveillance Question

Senator Kelly's framing cuts to the core of the mass surveillance issue: "Would you want them to have tools of AI to figure out whether you have a gun, to figure out what church you're going to?"

The technological capability to conduct mass surveillance using AI models is not hypothetical. Large language models can process surveillance data, correlate records across agencies, identify patterns in communication metadata, and assemble comprehensive profiles of individuals — automatically and at scale. The question is not whether the capability exists. It is whether legal and contractual restrictions prevent its deployment.

Anthropic argued that frontier AI systems could "assemble scattered data into comprehensive life pictures automatically and at massive scale," which is "incompatible with democratic values." The Pentagon argued this concern was "ideological."

The Fourth Amendment prohibits unreasonable searches. Whether AI-assisted mass surveillance constitutes an unreasonable search has not been litigated. The technology arrived before the law.

The Autonomous Weapons Question

Anthropic's second red line — no fully autonomous weapons without a human in the loop — rests on a technical argument as much as a moral one. Frontier AI systems, Anthropic argued, "lack sufficient reliability" for weapons that remove humans from targeting decisions, putting "America's warfighters and civilians at risk."

This is not an abstract concern. Israel's Lavender AI system generated 37,000 bombing targets in Gaza with reported approval times of 20 seconds — a pace that intelligence officers described as rubber-stamping AI recommendations rather than exercising independent judgment. The human was technically in the loop. The human was functionally a formality.

Anthropic's position was that if the human review is too fast to be meaningful, the system is autonomous in practice regardless of what the flowchart says. The Pentagon's position was that the distinction between "human-in-the-loop" and "human-on-the-loop" is a policy decision for the military, not a software restriction for a vendor.

For our analysis of autonomous weapons systems already in global deployment, see The Machines Are Already Built.

What Happens Now

The Legal Track

Anthropic's legal challenge to the supply-chain risk designation will test whether the government can use national security designations as punishment for refusing to modify commercial software. The outcome will establish precedent for every future negotiation between AI companies and government agencies.

If Anthropic wins, the government's leverage over AI companies narrows significantly. If Anthropic loses, the precedent extends beyond AI — any technology company whose products the government uses could face similar consequences for maintaining safety restrictions.

The Market Track

The $200 million Pentagon contract was a small fraction of Anthropic's estimated $14 billion in annual revenue. Amodei has noted that the company's valuation has only grown since taking its stand. Consumer and enterprise demand for Claude has not declined.

But the supply-chain risk designation creates secondary effects. Defense contractors must now choose between their Pentagon relationships and their Anthropic relationships. Companies in the defense industrial base — a vast ecosystem that extends far beyond weapons manufacturers into technology, logistics, and consulting — face a binary choice that the designation compels.

The Policy Track

No legislation has been introduced to establish clear rules for AI in military applications. The confrontation exposed a gap that Congress has acknowledged but not yet addressed: the United States has no statutory framework governing the relationship between AI companies' safety policies and government procurement demands.

Without legislation, future confrontations will follow the same pattern — ad hoc negotiations, ultimatums, executive actions, and court challenges. The result will be a patchwork of bilateral agreements with inconsistent terms, governed by the relative leverage of individual companies rather than coherent national policy.

The Safety Track

The RSP v3.0 changes remain in effect. Anthropic's founding commitment to halt development if capabilities outstripped safety measures has been replaced with a conditional commitment to delay development only if the company has a significant lead on competitors.

This means the company's most visible safety stand (military guardrails) was maintained at the cost of its most fundamental safety mechanism (the training pause). Whether this represents strategic prioritization or a net erosion of safety commitment depends on which guardrail you consider more important — and whether you believe the training pause was ever enforceable.

The Question That Remains

On February 26, as the deadline approached, Dario Amodei wrote: "These threats do not change our position: we cannot in good conscience accede to their request."

On February 27, Sam Altman wrote: "The DoW agrees with these principles, reflects them in law and policy, and we put them into our agreement."

The same principles. The same restrictions. One company blacklisted. The other welcomed.

The question is not whether Anthropic was right to maintain its guardrails. The OpenAI deal proves the guardrails were acceptable all along. The question is whether the cost Anthropic paid — and the message that cost sends to every other AI company — will make future acts of principle more likely or less.

In the short term, the answer is almost certainly less. The rational response to Anthropic's punishment is not to be more principled. It is to be more cooperative — to accept the same restrictions through quiet negotiation rather than public confrontation, to frame safety commitments as alignment with government policy rather than resistance to government demands.

The long-term answer depends on whether Anthropic's legal challenge succeeds, whether Congress acts, and whether the American public decides that the relationship between AI companies and state power requires guardrails that neither side can unilaterally remove.

Dario Amodei bet his company's government relationships that the answer is yes. The next twelve months will determine whether he was right.

---

For the full timeline of the Pentagon's autonomous weapons procurement strategy, see The Skynet Procurement Timeline. For our earlier analysis written before the blacklisting, see The Anthropic Ultimatum. For Anthropic's broader impact on the enterprise software market, see The $130 Billion Month.