Pentagon Inks AI Deals with 7 Tech Giants for Classified Networks, Sidelines Anthropic
The U.S. Department of Defense has formalized classified-network AI agreements with OpenAI, Google, Nvidia, Microsoft, Amazon, SpaceX’s xAI, and Reflection AI, openly excluding the one company that refused to strip its safety guardrails.
On May 1, 2026, the U.S. Department of Defense announced it had secured AI agreements with seven leading technology companies, granting their models access to Impact Level 6 and 7 classified networks covering everything from intelligence analysis to weapons targeting. One name was conspicuously absent: Anthropic, maker of the Claude models that, until recently, held the only frontier AI authorization on those same networks.
The exclusion didn’t come quietly. It followed a two-month standoff over what the Pentagon demanded and what Anthropic refused to accept: the removal of contractual safeguards against using AI for autonomous kill decisions and mass domestic surveillance of American citizens. When negotiations collapsed in February, the DoD took the extraordinary step of designating Anthropic a “supply-chain risk”, a label typically reserved for foreign adversaries like Huawei.
The announcement marks a decisive turn in how the U.S. military intends to field AI in warfighting operations. Seven companies have now agreed, in writing, to provide access for what DoD contracts describe as “any lawful governmental purpose.” The question of what that phrase actually permits, and who decides, sits at the center of a federal lawsuit, a temporary court injunction, and a growing split inside the AI industry itself.
The Seven Companies and What They’re Providing
The agreements cover AI deployments on the Pentagon’s most sensitive networks. Impact Level 6 handles secret-classified data, operational planning, intelligence feeds, logistics modeling. Impact Level 7 reaches into top-secret territory: mission-critical command and control, weapons targeting, and battlefield data fusion. The companies now authorized at those levels are:
OpenAI
GPT series models, including agentic capabilities for autonomous task execution across classified pipelines.
Gemini models, building on a prior $200M baseline contract signed April 28. Google signed a separate classified deal first among the seven.
xAI (SpaceX)
Grok models, providing Elon Musk’s frontier AI into the DoD’s core decision-support stack.
Nvidia
AI infrastructure and chips, the hardware backbone underpinning inference at classified classification levels.
Microsoft + AWS
Azure AI and Copilot alongside Amazon Web Services cloud AI services, both already entrenched DoD cloud providers.
Reflection AI
A frontier-model startup earning its first major government contract, a signal that DoD is deliberately seeding competition beyond established players.
Together, these companies represent a combined agentic AI contract valued at roughly $800 million across four of the parties, with each major provider receiving approximately $200 million in agentic AI contract awards. The GenAI.mil platform, the Pentagon’s internal AI access system, already had 1.3 million DoD personnel generating tens of millions of prompts and deploying hundreds of thousands of AI agents within its first five months of operation.
GenAI.mil by the numbers (first 5 months): 1.3 million DoD personnel onboarded, tens of millions of prompts processed, hundreds of thousands of autonomous agents deployed. The platform now expands to Impact Level 6 and 7 networks with all seven vendors above.
How Anthropic Got Blacklisted — and Why It Matters
Until early 2026, Anthropic held a uniquely privileged position. Claude was the only frontier large language model formally authorized to operate on classified DoD networks, integrated into Palantir’s Maven Smart System, the AI platform that supported Pentagon operations in Iran. That changed when Secretary of Defense Pete Hegseth issued a January 9 memorandum requiring all DoD AI contracts to include “any lawful use” language within 180 days.
“The Pentagon would not employ AI models that won’t allow you to fight wars.”
Pete Hegseth, Secretary of Defense, February 2026
Anthropic’s position, as stated by CEO Dario Amodei during negotiations, was that the AI model should be used in accordance with what it can “reliably and responsibly do.” The company insisted on maintaining two specific contractual safeguards: a prohibition on using Claude for autonomous weapons systems without human-in-the-loop oversight, and a ban on mass domestic surveillance of U.S. citizens. The Pentagon rejected both conditions.
Negotiations collapsed in February. On March 5, the DoD formally designated Anthropic a “supply-chain risk”, an unprecedented move against a domestic AI company. The label carries practical teeth: it bars military agencies and their contractors from using Anthropic’s products. The designation normally applies to foreign-linked technology suppliers like telecommunications hardware from companies with ties to China’s government.
Precedent alert: A “supply-chain risk” designation against a U.S. AI company is without modern precedent. The legal authority used derives from the same statutes applied to Huawei and ZTE. Anthropic’s legal team argues this represents an unconstitutional use of national security emergency powers against a domestic firm for refusing to weaken its ethical policies.
The other six companies took a different approach. OpenAI reportedly proposed a separate technical safety stack while contractually deferring all usage decisions to existing U.S. law. Google agreed to the “any lawful governmental purpose” framing despite internal objections. As DeepMind research scientist Alex Turner noted in late April, that framing gives Google no practical veto over how the Pentagon deploys its models.
“Google can’t veto usage, the reliance on aspirational language without any legal constraints is the core problem here.”
Alex Turner, Research Scientist, DeepMind, April 29, 2026
Inside the Classified Networks: What These AI Systems Actually Do
Impact Level 6 and 7 aren’t abstract categories. They define the security architecture, vetting requirements, and permissible use cases for everything running on those networks. Below is what the DoD’s own technical framework requires at each tier.
| Classification Level | Security Standard | Primary Use Cases | AI Applications |
|---|---|---|---|
| Impact Level 6 (Secret) | FedRAMP High + DoD IL6 authorization | Intelligence analysis, operational planning, ISR data fusion | Data synthesis, situational awareness, logistics optimization |
| Impact Level 7 (Top Secret) | Highest clearance level, continuous monitoring | Weapons targeting, mission-critical C2, strategic planning | AI-assisted targeting, predictive battlefield modeling, autonomous agent deployment |
The DoD’s stated objectives for these integrations are “streamlining data synthesis, elevating situational understanding, and augmenting warfighter decision-making.” In practice, that means AI models processing classified intelligence feeds in near-real time, generating targeting recommendations, and managing logistics chains that span multiple theaters simultaneously. Hundreds of thousands of AI agents are already operating autonomously within the broader GenAI.mil infrastructure.
“The Pentagon wants to go beyond last year’s limits on autonomous weapons and expand AI from intelligence and reconnaissance to kinetic uses, such as selecting and engaging targets with drones.”
Vanessa Vos, Researcher, Bundeswehr University Munich, March 4, 2026
All vendors must meet FedRAMP High certification and comply with a zero-trust architecture mandate that runs through September 2027. They also operate under DoD Directive 3000.09, the autonomous weapons policy, which the Secretary of Defense can adjust without congressional approval. That last point is critical: the policy guardrails governing how these AI systems engage with targeting decisions sit entirely within the executive branch’s discretion.
The Staff Reluctance Problem
There’s a wrinkle the Pentagon’s announcement didn’t address. Multiple reports indicate that DoD staff who routinely used Claude for classified work are reluctant to switch. Claude’s capabilities in complex reasoning and nuanced synthesis earned it a strong internal following. Replacing it with models that staff consider inferior, at least for certain analytical tasks, creates uneven capability across units. That’s not a hypothetical concern; it’s an operational risk the DoD is absorbing as the price of its policy choice.
The Financial Stakes: $380 Billion in the Balance
For Anthropic, this isn’t just a policy dispute. It’s an existential financial threat. The company’s pre-blacklist market valuation stood at approximately $380 billion, according to analysis published April 30. The direct contract loss is quantifiable: the DoD deal under negotiation was worth up to $200 million, part of an $800 million agentic AI contract shared across four providers. The indirect damage is harder to measure but potentially far larger.
| Stakeholder | Financial Exposure | Direction |
|---|---|---|
| Anthropic | $200M direct contract loss; billions in 2026 enterprise revenue at risk; $380B valuation under pressure | Negative |
| OpenAI | ~$200M agentic AI contract; expanded defense pipeline access | Positive |
| $200M+ (expanded from prior baseline contract); classified network access for Gemini | Positive | |
| Nvidia | Infrastructure revenue across all seven vendor deployments; chip demand tied to IL6/7 inference | Strongly Positive |
| Palantir | $10B+ Army data contracts; $795M+ Maven Smart System support — now runs on rival models | Mixed |
| Reflection AI | First major government contract; instant defense-sector credibility | Strongly Positive |
| Anduril | $20B Lattice AI C2 Enterprise contract (Army); aligned with DoD’s kinetic AI direction | Positive |
Anthropic’s legal filings describe the revenue impact as running into “multiple billions” during 2026 alone, according to analysis by Pearl Cohen published March 25. An IPO that had been in preparation becomes significantly more complicated when the company is formally designated a risk to national security supply chains. Enterprise customers in adjacent government and contractor markets face their own compliance questions about continuing to use Claude.
The Lawsuit That Temporarily Stopped the Clock
Anthropic didn’t accept the blacklist quietly. On March 9, the company filed two simultaneous federal lawsuits: one in the Northern District of California and a second in the D.C. Circuit Court of Appeals. The legal theory combined First Amendment arguments, that the government can’t penalize a company for the speech embedded in its AI policies, with administrative law claims that the DoD exceeded its statutory authority.
On March 26, a federal judge granted a temporary stay of the “supply-chain risk” designation, pausing its enforcement while the litigation proceeds. That stay doesn’t reinstate Anthropic’s contracts. It doesn’t undo the May 1 announcement. It means the legal classification remains contested while the deals move forward with the other seven vendors.
The case raises questions with no clean precedent. Can the government compel an AI company to remove ethical constraints as a condition of federal contracting? Does a “supply-chain risk” designation require evidence of actual security risk, or can it rest on policy disagreement? And if companies can be blacklisted for maintaining safety guardrails, what incentive structure does that create across the industry?
“Statements outside formal AI contracts do not alter legal liability if ethical or legal concerns arise later.”
Tuncer, Legal Expert, Anadolu Agency, March 1, 2026
Congress has started paying attention. Axios reported that several lawmakers are exploring legislation to establish minimum guardrails for military AI deployments, a direct response to the Anthropic dispute. Any such legislation would face the same executive-branch resistance that produced the original standoff.
Safety vs. Speed: A Race the Industry Can’t Ignore
Step back from the specific contracts and what emerges is a structural incentive problem. The Pentagon has now demonstrated that companies maintaining strong internal safety policies on autonomous weapons and surveillance can be shut out of the defense market entirely. Companies that defer those decisions to existing law, and accept that the executive branch will define what that law permits, get access to some of the largest government contracts available.
“Race to the bottom where the most compliant firms win”, on Pentagon blacklisting dynamics.
Geoffrey Gertz, Independent Defense AI Analyst, February 16, 2026
The AI industry’s internal debate over this isn’t theoretical. Some researchers argue that companies without government contracts lose the ability to shape how AI is deployed in high-stakes settings. Others contend that accepting “any lawful use” language, where “lawful” is defined unilaterally by the government using the AI, represents a fundamental abdication of responsibility.
“US military’s reliance on fluid domestic definitions due to lack of international law creates legal loopholes for mass surveillance and autonomous weapons use.”
Firdevs Bulut Kartal, Author, Anadolu Agency, March 2, 2026
The international dimension compounds the problem. The International Committee of the Red Cross and several allied governments have pushed for binding treaties governing autonomous weapons. The U.S. now has seven major AI vendors operating on classified military networks under contracts that explicitly reject company-level ethical constraints, and no international legal framework that would fill the gap.
- DoD Directive 3000.09 governs autonomous weapons policy and can be modified by the Secretary of Defense without congressional approval
- None of the seven vendor agreements include third-party audit rights or external oversight mechanisms
- The “any lawful use” framing places the entire interpretive burden on the executive branch
- No allied nation has adopted an equivalent “AI-first warfighting force” doctrine at this speed or scale
- Zero-trust architecture (mandatory by September 2027) addresses cybersecurity, not policy compliance
For the vendors themselves, the tension isn’t abstract. Both Google and OpenAI faced significant internal employee pushback over prior military AI work. Both have now signed contracts that their own researchers publicly criticize. The question isn’t whether that tension exists, it’s whether it produces any meaningful constraint on deployment decisions.
Frequently Asked Questions
Why was Anthropic excluded from Pentagon AI deals?
Anthropic refused to remove two contractual safeguards, one prohibiting autonomous weapons use without human oversight, and one banning mass domestic surveillance, that the Pentagon required all vendors to drop. When negotiations failed in February 2026, the DoD designated Anthropic a “supply-chain risk,” barring military use of its models.
What does “Impact Level 6 and 7” mean for military AI?
Impact Level 6 covers secret-classified networks used for intelligence analysis and operational planning. Impact Level 7 is top-secret, covering weapons targeting and mission-critical command and control. Both require FedRAMP High certification and continuous security monitoring.
What is the “any lawful use” clause in DoD AI contracts?
It’s a contract provision, mandated by Secretary Hegseth’s January 2026 memo, requiring AI vendors to permit any use the government considers lawful. Critics argue it gives vendors no ability to restrict how their models are deployed for autonomous weapons or surveillance, with the government as the sole arbiter of what’s permitted.
Has Anthropic’s lawsuit succeeded in blocking the blacklist?
A federal judge issued a temporary stay of the “supply-chain risk” designation on March 26, 2026, pausing enforcement while litigation proceeds. However, the stay didn’t restore Anthropic’s contracts, and the Pentagon’s May 1 deals with seven other companies moved forward regardless.
Which companies signed Pentagon classified AI deals in May 2026?
Seven companies: OpenAI, Google, Nvidia, Microsoft, Amazon Web Services, xAI (SpaceX’s AI division, providing Grok), and Reflection AI, a frontier-model startup receiving its first major government contract. Anthropic was explicitly excluded.
How large is the Pentagon’s AI investment across these deals?
The agentic AI contracts for four of the seven companies total approximately $800 million, with each receiving around $200 million. Broader defense AI context includes a $20 billion Anduril Lattice contract, $10 billion-plus Palantir Army contracts, and a $9 billion Joint Warfighting Cloud Capability ceiling.
What is GenAI.mil and how widely is it used?
GenAI.mil is the Pentagon’s official AI access platform for DoD personnel. Within its first five months it onboarded 1.3 million military personnel, processed tens of millions of prompts, and deployed hundreds of thousands of autonomous AI agents across various operational tasks.
What are the cybersecurity requirements for these AI deployments?
All vendors must meet FedRAMP High certification and Impact Level 6 or 7 authorization. The DoD has also mandated zero-trust architecture across its AI deployments, with a compliance deadline of September 2027. Zero trust governs network access controls but doesn’t address policy compliance or autonomous weapons constraints.
What Comes Next in Military AI
The Pentagon’s May 1 announcement is less a conclusion than a line drawn in the sand. Seven companies now hold classified-network access under contracts that prioritize deployment speed over independent safety oversight. One company is fighting that framework in federal court while watching its valuation erode. And the broader AI industry is absorbing the lesson: in the defense market, safety constraints are a liability, not a selling point.
The short-term winners are obvious. OpenAI, Google, and Nvidia gain enormous revenue and strategic positioning. Reflection AI graduates from startup to defense contractor overnight. The long-term picture is murkier. If autonomous AI targeting systems fail in the field, or if domestic surveillance applications produce a political crisis, the companies that signed “any lawful use” agreements will find those contracts suddenly very visible. The absence of contractual accountability doesn’t eliminate operational accountability. It just shifts when it arrives.
For the broader AI safety community, the Anthropic case establishes a troubling precedent: a domestic AI company can be designated a national security risk not for building dangerous technology, but for refusing to make its technology less safe. Whether Congress, the courts, or allied governments move to address that precedent will define the regulatory environment for military AI for the decade ahead.