Pentagon Seeks AI That Rarely Refuses Orders, Raising Concerns About Autonomous Weapons

The U.S. Department of Defense sought a custom version of OpenAI's technology engineered to refuse military commands as infrequently as possible, according to documents obtained through a Freedom of Information Act lawsuit. The revelation, published September 8, 2026, exposes a fault line running through every major AI company's Pentagon relationship: the gap between commercial safety architecture and military operational requirements.

Documents released to The Intercept as part of a FOIA lawsuit describe a contract clause calling for "OpenAI Mission Models" built for "national security use cases" and featuring "minimal refusal rates." The contract in question, labeled version "P00003," represents an expansion of an earlier prototype deal valued at up to $200 million over two years. A separate document signed by both OpenAI and the federal government on February 6 indicates OpenAI agreed to the expanded terms on that date. OpenAI spokesperson Nate Evans disputed the characterization, stating the company never agreed to "minimal refusal rates" language and that the document The Intercept received was a draft proposal the company rejected. The Pentagon's own lawyers initially confirmed the document as the signed, executed contract, then retracted that confirmation hours later, after The Intercept contacted OpenAI for comment .

Trevor Tiedeman, a special assistant to the under secretary of war for research and engineering and a former staffer on President Donald Trump's reelection campaign, told The Intercept there may be "stray voltage or wires crossing" between the FOIA materials, executed contracts, and draft versions. He said he would investigate. He subsequently stopped responding. Department of Defense spokesperson Jacob Bliss later stated that the phrase "minimal refusal rates" does not appear in any active Department of War contract with OpenAI .

For institutional investors, the contradictions are not a bureaucratic footnote. They are a pricing signal. The credibility of AI companies' safety commitments, their contractual relationships with the federal government, and their exposure to regulatory and reputational risk all feed directly into valuation. When the government's own lawyers cannot confirm which version of a contract is operative, the governance structure around these deals is weaker than the market has assumed.

A $200 Million Deal and the Guardrail Problem

The Pentagon contract at the center of this story carries a ceiling value of up to $200 million over two years, according to The Intercept . That figure covers a deal that began as a prototype agreement in 2025 and expanded in scope in 2026. OpenAI, Google, xAI, and Anthropic all agreed in 2025 to develop militarized prototypes of their frontier AI systems for military applications including logistics, intelligence decision-making, and what the documents describe as general "warfighting."

The deliverables described in the contract document include "Testing, Evaluation, and Refinement of OpenAI Mission Models" for "national security problem sets." The document defines OpenAI Mission Models specifically as models "designed for national security use cases" with "minimal refusal rates." No further description of what those national security problem sets entail appears in the released materials .

The commercial version of ChatGPT, when asked to help prioritize drone-based airstrike targets, returns a refusal. OpenAI and its competitors prohibit public users from applying their models to weapons of mass destruction development. A model engineered to minimize those refusals, deployed on classified networks, would represent a fundamentally different product than anything these companies currently sell commercially .

The Anthropic Precedent and the Guardrail Negotiation Playbook

OpenAI is not the only company navigating this tension. The Intercept reported that earlier in 2026, the Pentagon sought to extend its AI agreements to classified computer networks, triggering a direct confrontation with Anthropic over the company's ability to constrain military applications of its technology . That standoff is now playing out in a second arena: on September 8, 2026, the same day The Intercept published the OpenAI FOIA findings, The Verge reported that a class-action lawsuit had been filed against Anthropic by subscribers of its Max plan, priced at $100 to $200 per month, who allege the company misled them about what the tier delivered .

The simultaneity is not coincidental. It reflects a single structural problem: AI companies built their consumer and enterprise products around safety guardrails and usage promises. Military and power-user contracts stress-test both. When guardrails are loosened for the Pentagon and usage limits disappoint paying customers, the same underlying question surfaces: what exactly did you agree to sell, and to whom?

For PE and institutional investors assessing AI exposure, the Anthropic litigation adds a concrete liability vector to a company that has raised at a multi-billion dollar valuation. Terms of that litigation were not fully disclosed in available sources, but the plaintiff class consists of Max plan subscribers paying between $100 and $200 per month who argue the product did not match representations .

Microsoft Moves to Capture the Trust Deficit

While OpenAI and Anthropic absorb scrutiny over military contracts and consumer lawsuits, Microsoft executed a different kind of institutional play. On September 9, 2026, one day after the Pentagon FOIA story broke, The Verge reported that Microsoft had reached an agreement with the American Federation of Teachers, the second-largest teachers union in the United States, and its New York City affiliate the United Federation of Teachers .

The agreement commits Microsoft to ten contractually enforceable principles for AI use in schools. These include a pledge not to train AI models on student or educator data, limits on data collection, plain-language disclosure to families, a prohibition on AI companions, and mandatory human review for high-risk decisions. The agreement came one week after New York City and Los Angeles both imposed one-year bans on AI tools for many students .

The strategic logic is direct. Microsoft is moving into the institutional trust gap that OpenAI's Pentagon entanglement and Anthropic's litigation have created. School districts, state governments, and regulated institutions evaluating AI adoption now have a contractually binding framework from Microsoft that neither OpenAI nor Anthropic has matched in comparable institutional settings. That positioning has revenue implications that extend well beyond the education vertical.

The Governance Discount in AI Valuations

Institutional capital pricing AI companies in 2026 has largely applied a growth premium to revenue trajectories without fully discounting governance risk. The Pentagon FOIA dispute changes that calculus.

The core issue is contract opacity. The DoD and OpenAI are presenting contradictory accounts of what terms are operative on a deal worth up to $200 million. The government's own lawyers reversed their confirmation within hours. A named senior Pentagon official, Tiedeman, described the situation as "stray voltage" and then went silent. Neither OpenAI nor the Pentagon provided the allegedly correct final contract .

For any institutional investor, this opacity on a disclosed contract raises questions about what undisclosed terms look like on classified network deployments. The Pentagon's earlier push to extend these AI agreements across classified systems implies a much larger procurement pipeline than the $200 million prototype ceiling suggests. The gap between the public-facing safety architecture and the military-use product is not theoretical. It is now documented.

The Microsoft-AFT agreement illustrates the counterfactual: governance commitments can be made explicit, contractual, and enforceable. The companies that build that framework for institutional clients, governments, and regulated industries will price at a premium to those that do not.

The Plocamium View

The market is treating the Pentagon FOIA story as a reputational event for OpenAI. That framing is too narrow.

What the documents reveal is that the frontier AI companies entered 2025 and 2026 defense contracts without resolving the fundamental tension between their safety architectures and military operational requirements. The "minimal refusal rates" clause, whether or not it survived to the final contract, describes something the Pentagon wanted and believed it could obtain. That demand does not disappear because the language was removed from one document.

The second-order effect is regulatory. Congress and international bodies evaluating AI governance now have a concrete data point: the U.S. military sought an AI with systematically reduced safety refusals, and at minimum one draft contract reflected that goal. That is exactly the kind of evidence that accelerates formal AI oversight legislation. The EU AI Act framework, which classifies high-risk AI applications with specific compliance requirements, was already designed with military-adjacent use cases in mind. A documented Pentagon effort to strip guardrails from a commercial LLM for national security applications will feature in every subsequent regulatory proceeding.

For PE investors, the actionable insight is a bifurcation trade. Companies with explicit, enforceable governance frameworks targeting institutional buyers, as Microsoft is building, will trade at a different multiple than those navigating open-ended government contract disputes. The Anthropic class-action adds a consumer litigation overlay that compounds the governance discount on that company specifically.

The defense AI market is real, growing, and strategically important. But the companies that capture durable margin in that market will be those that resolve the guardrail architecture question in writing, with enforced contractual terms, rather than those that leave it to be discovered through FOIA lawsuits. The Pentagon's demand for minimal refusals is a product specification. The companies that answer it with a coherent governance structure, rather than contradictory statements, will own the procurement pipeline.

The Bottom Line

The Pentagon's documented pursuit of an AI system engineered to minimize refusals is not primarily a story about one contract clause. It is a stress test of the entire AI defense procurement framework, and the companies that have not resolved the tension between commercial safety guardrails and military use cases face both regulatory and reputational exposure that is not yet priced into their valuations. Microsoft's move to lock in contractually enforceable AI governance with the AFT signals that institutional trust is becoming a competitive moat. The next twelve months will separate the AI companies that formalize their governance architecture from those that litigate it after the fact.