When the world's most safety-conscious AI company faced the Defense Production Act, the age of voluntary AI ethics ended
Executive Summary
- The Pentagon has given Anthropic until 5:01 PM Friday, February 27 to grant unrestricted military access to Claude AI—or face forced compliance via the Defense Production Act, a Cold War-era law never before wielded against an AI company.
- Elon Musk's xAI has already agreed to classified military use of Grok, while OpenAI and Google are reportedly "close" to full compliance—leaving Anthropic as the sole holdout in a corporate compliance cascade.
- The crisis reveals a fundamental power shift: the question of who controls AI's ethical boundaries—Silicon Valley or the Pentagon—has been answered, and the answer is not the one the AI safety community hoped for.
Chapter 1: The Ultimatum
At a tense meeting on Tuesday, February 25, Defense Secretary Pete Hegseth delivered a message to Anthropic CEO Dario Amodei that was remarkable for both its bluntness and its historical weight. According to a senior Pentagon official who spoke to NBC News, Hegseth told Amodei: allow Claude to be used for "all lawful military purposes" by 5:01 PM Friday, or the Pentagon would invoke the Defense Production Act to compel compliance.
"Anthropic has until 5:01pm Friday to get on board with the Department of War," the Pentagon official stated. "If they don't get on board, the Secretary of War will ensure the Defense Production Act is invoked on Anthropic, compelling them to be used by the Pentagon."
The DPA, signed by President Truman in 1950 during the Korean War, grants the executive branch extraordinary powers to direct private companies to prioritize production critical to national security. It was invoked during the COVID pandemic to force production of medical supplies. But using it against an artificial intelligence company—to override its voluntary ethical restrictions—would be unprecedented in American history.
The Pentagon also threatened a second punishment: designating Anthropic as a "supply chain risk," the same classification applied to Chinese firms like Huawei. Such a designation would effectively blacklist Anthropic from the entire defense ecosystem, forcing every military contractor to sever ties with the company.
The irony is thick. Anthropic was, until recently, the only AI company cleared for use by the Pentagon. Its Claude model was deployed through a partnership with Palantir in classified settings. Now, the company's very insistence on maintaining safety guardrails has made it a target.
Chapter 2: The Venezuela Trigger
The fracture between Anthropic and the Pentagon can be traced to a specific event: the January 3 capture of Venezuelan President Nicolás Maduro during Operation Absolute Resolve.
According to reporting by The Hill, the Pentagon used Anthropic's Claude model—accessed through Palantir's defense platform—during the operation. When Anthropic learned of this, the company reached out to Palantir to ask how its AI had been employed in the military raid. Palantir, in turn, flagged this inquiry to the Pentagon.
For the Defense Department, Anthropic's question was a red flag. It suggested the AI company viewed itself as having oversight authority over how the military used tools it had already purchased. In the Pentagon's view, this was unacceptable: once the military acquires a capability, its deployment is a sovereign decision, not subject to a vendor's ethical review.
For Anthropic, the concern was existential. The company was founded in 2021 by former OpenAI researchers specifically to build AI that was safe and controllable. Its Responsible Scaling Policy—the corporate equivalent of a Hippocratic oath—explicitly prohibited the use of Claude for "fully autonomous weapons that operate without human involvement" and "mass surveillance of domestic populations."
The Venezuela incident crystallized a question that had been theoretical until that moment: What happens when a government decides that an AI company's ethical boundaries are obstacles to national security?
Chapter 3: The Compliance Cascade
Anthropic is not alone in facing Pentagon pressure—but it is alone in resisting.
On February 23, Elon Musk's xAI signed a deal allowing the Pentagon to use Grok in classified settings. The agreement was notable not just for its content but for its speed: xAI completed negotiations in a matter of days, with no public discussion of ethical boundaries or use restrictions.
A Pentagon official confirmed to Politico that OpenAI and Google are "close" to similar agreements. Neither company has publicly articulated red lines comparable to Anthropic's.
This creates a compliance cascade—a dynamic where each company's capitulation increases the pressure on holdouts. When xAI agreed, the Pentagon's leverage over Anthropic increased: the Defense Department could credibly argue it had alternatives, reducing Anthropic's bargaining power while simultaneously demonstrating that resistance was futile.
The $200 million in defense contracts at stake is almost beside the point. The real question is whether any major AI company can maintain independent ethical standards when the world's largest military customer demands otherwise.
| Company | Pentagon Status | Autonomous Weapons | Mass Surveillance | Classified Use |
|---|---|---|---|---|
| xAI (Grok) | ✅ Full compliance | No restrictions stated | No restrictions stated | Agreed Feb 23 |
| OpenAI | 🟡 "Close" to deal | Previously cautious | Previously cautious | Negotiating |
| 🟡 "Close" to deal | Withdrew from Maven (2018), reversed | Previously cautious | Negotiating | |
| Anthropic | ❌ Holdout | Explicit prohibition | Explicit prohibition | Offered missile defense only |
| Meta (Llama) | ✅ Open-source | No restrictions (open model) | No restrictions | N/A (open) |
Chapter 4: The Compromise That Wasn't Enough
Anthropic did not simply refuse the Pentagon's demands. In December 2025, the company updated its contract language to explicitly allow Claude to be used for missile defense and cyber defense—two areas where the case for AI assistance is most compelling and where the ethical objections are weakest.
"Every iteration of our proposed contract language would enable our models to support missile defense and similar uses," an Anthropic spokesperson told NBC News.
But this compromise was rejected. During negotiations, Pentagon officials—led by Undersecretary of Defense Emil Michael—presented hypothetical scenarios, including an ICBM launched at the United States. When Pentagon officials asked whether Anthropic's guardrails might somehow delay a U.S. response, Anthropic officials said they could lift restrictions in such scenarios. But Pentagon leadership was "not fully satisfied" and "did not want to be beholden to the private company."
The word "beholden" is key. The Pentagon's objection is not primarily about capability—it is about authority. The military does not want to negotiate with a vendor over the terms of use during a crisis. It wants unconditional access, available at all times, for all lawful purposes.
This is, in many ways, a reasonable position for a defense establishment. It is also a position that renders voluntary AI safety commitments meaningless—because the moment a guardrail becomes inconvenient, the government can simply order it removed.
Chapter 5: The Safety Surrender
The most telling development may not be the Pentagon's ultimatum but Anthropic's own response. On Tuesday—the same day Hegseth delivered his ultimatum—Anthropic released version 3 of its Responsible Scaling Policy.
The updated policy acknowledged what its authors described as a shift in the "policy environment." The document stated: "The policy environment has shifted toward prioritizing AI competitiveness and economic growth, while safety-oriented discussions have yet to gain meaningful traction at the federal level."
This is a remarkable admission from a company that built its entire brand on the premise that AI safety should come first. The RSP v3 update loosens several commitments, reflecting the commercial and political reality that safety-first companies face existential disadvantages when competitors accept fewer restrictions.
CEO Dario Amodei had already signaled this shift in a February 17 interview, suggesting the company might need to rebalance its approach to remain competitive. The "woke AI" label—deployed relentlessly by Hegseth, who in a January speech at SpaceX declared "Department of War AI will not be woke"—has turned safety commitments into a political liability.
The trajectory is clear. In 2023, Anthropic published the gold standard of AI safety frameworks. In 2024, it secured Pentagon contracts while maintaining guardrails. In 2025, it lost its exclusive position as the Pentagon's preferred AI partner. In 2026, it faces a choice between capitulation and blacklisting.
Chapter 6: Scenario Analysis
Scenario A: Anthropic Complies by 5:01 PM (55%)
Rationale: The company has already begun loosening its RSP. The $200 million contract, the threat of supply chain blacklisting, and the competitive pressure from xAI/OpenAI/Google create overwhelming incentives to comply. Anthropic may issue a carefully worded statement reframing compliance as "responsible engagement."
Historical precedent: Google abandoned its "Don't Be Evil" principles regarding Project Maven in 2018—then quietly reversed course and re-entered defense AI. Corporate ethical stances rarely survive sustained government pressure. During the Cold War, defense contractors routinely abandoned voluntary restrictions when faced with government demands.
Trigger conditions: Board-level decision prioritizing commercial survival over safety branding. Likely accompanied by face-saving language about "expanded responsible use."
Investment implications: Anthropic enterprise valuation stabilizes; Palantir partnership deepens. Defense AI market consolidates around 4-5 providers with uniform terms.
Scenario B: DPA Invocation (30%)
Rationale: If Anthropic refuses or attempts to negotiate an extension, Hegseth has publicly committed to DPA invocation. The political incentive to follow through is high: backing down would undermine the administration's "woke AI" narrative and embolden other companies to resist.
Historical precedent: The DPA was invoked 18 times during COVID for medical supplies. However, using it against a technology company to override ethical restrictions would be qualitatively different—closer to the government ordering a newspaper to publish certain content than ordering a factory to produce ventilators.
Trigger conditions: Anthropic submits a compromise that falls short of "all lawful purposes." Hegseth, politically committed, invokes DPA over the weekend. Legal challenges follow within days.
Legal vulnerability: The DPA requires a presidential finding that the product is essential to national defense. Anthropic could challenge whether AI chatbot access constitutes a "product" under DPA's statutory framework, which was written for physical goods. This could reach federal courts within weeks.
Investment implications: AI sector volatility spikes. Legal uncertainty creates regulatory risk premium across the sector. Defense contractors face supply chain disruption.
Scenario C: Negotiated Extension (15%)
Rationale: Both sides have incentives to avoid a public rupture. A quiet 30-60 day extension, with incremental concessions from Anthropic, would allow face-saving on both sides.
Trigger conditions: Back-channel intervention by White House officials or congressional allies who view DPA invocation as legally risky ahead of IEEPA Supreme Court complications. Anthropic agrees to expanded categories while maintaining nominal guardrails.
Investment implications: Uncertainty persists but is contained. AI safety as a brand differentiator continues to erode.
Chapter 7: Investment Implications
The Pentagon's ultimatum is not just an Anthropic story—it is a repricing event for the entire AI governance landscape.
Winners:
- Palantir (PLTR): Already the Pentagon's preferred integration layer for AI. Full Anthropic compliance deepens the partnership; DPA invocation routes even more work through Palantir as the trusted intermediary.
- Defense primes (LMT, RTX, NOC): AI integration accelerates across weapons systems as ethical friction disappears.
- xAI/SpaceX: First-mover advantage in classified AI sets up the $1.5 trillion IPO narrative.
Losers:
- Anthropic's private valuation ($380B): Any outcome damages either its safety brand or its government revenue. The "premium for responsibility" thesis collapses.
- AI safety ecosystem: Academic research groups, safety-focused startups, and policy organizations lose their most prominent corporate champion.
- European AI companies: EU AI Act's strict safety requirements become a competitive disadvantage if U.S. companies are now government-compelled to prioritize capability over safety.
Structural shift: The era of voluntary AI governance is ending. The question is no longer whether AI companies will self-regulate but whether governments will permit them to. The DPA threat establishes a template: any future administration can use national security law to override corporate AI ethics. This fundamentally changes the risk calculus for AI investment, favoring companies that align early with government priorities over those that maintain independent ethical frameworks.
Conclusion
The Defense Production Act was born in 1950, when the United States needed steel mills to produce armor plate for Korean War tanks. Seventy-six years later, it may be used to force a San Francisco startup to remove safety guardrails from an artificial intelligence system.
The symmetry is imperfect but revealing. In both cases, the government's argument is the same: national security overrides private prerogatives. The difference is that in 1950, the government was ordering companies to produce more. In 2026, it is ordering a company to restrict less—to remove the very safeguards the company was designed to uphold.
Whether Anthropic complies, resists, or negotiates, the precedent is already set. The Pentagon has demonstrated that it possesses both the legal tools and the political will to override AI safety commitments. The Defense Production Act need never actually be invoked for its threat to reshape the industry. Every AI company CEO now knows the cost of maintaining guardrails the government finds inconvenient.
Friday at 5:01 PM is not just a deadline for Anthropic. It is the moment when the age of voluntary AI ethics officially ends—and the age of state-directed AI begins.
Leave a Reply