The Pentagon designated Anthropic a supply-chain risk and barred its use
In a move that reflects both the promise and the peril of artificial intelligence in the service of national power, the Pentagon has formalized partnerships with seven major AI firms — including OpenAI, Google, and Microsoft — granting them access to the military's most classified network environments. Conspicuously absent is Anthropic, whose insistence on ethical guardrails and whose Mythos model's cyber capabilities have placed it at odds with a Defense Department that prizes operational freedom above constraint. The episode illuminates a deeper tension of our technological moment: the question of who sets the terms when the tools of intelligence meet the instruments of war.
- The Pentagon has quietly but decisively expanded its AI infrastructure, giving seven companies access to classified networks used by over 1.3 million Defense Department personnel — a scale that signals AI is no longer peripheral to military operations, but central.
- Anthropic's absence from the list is not an oversight — the company was formally designated a supply-chain risk, a rare and pointed rebuke that effectively bars its tools from the entire Defense Department contractor network.
- The flashpoint is Mythos, Anthropic's cyber-capable AI model, which alarmed Pentagon officials enough that the CTO called it 'a separate national security moment' — even as the broader dispute over guardrails remains unresolved.
- President Trump's offhand remark that Anthropic is 'shaping up' has introduced ambiguity into what appeared to be a firm exclusion, leaving open the possibility of a negotiated return if both sides can find common ground on terms of use.
- The Pentagon's deliberate spread across seven vendors signals a strategic doctrine of diversification — reducing the risk of over-reliance on any single AI partner while pressing forward with capabilities that carry their own profound risks.
The Pentagon announced Friday that it had reached agreements with seven AI companies — SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, and Amazon Web Services — to integrate their systems into its most sensitive classified networks. The decision reflects a deliberate strategy of vendor diversification, ensuring that no single company holds outsized influence over the military's growing AI infrastructure. Over the past five months alone, more than 1.3 million Defense Department personnel have accessed GenAI.mil, the department's primary AI platform, underscoring how rapidly these tools have moved from experiment to operational reality.
The announcement's most telling detail, however, is an absence. Anthropic — the company behind Claude — was not included. Earlier this year, the Pentagon formally designated Anthropic a supply-chain risk, barring its use across the department and its contractor network. The underlying conflict is philosophical as much as technical: Anthropic has pushed for guardrails on how its systems can be used in military contexts, while the Pentagon has resisted what it sees as unacceptable limits on its operational latitude.
The dispute sharpened with the emergence of Mythos, Anthropic's AI model built with advanced cyber capabilities. Pentagon CTO Emil Michael described it as 'a separate national security moment,' reflecting genuine alarm about a system capable of amplifying offensive hacking operations. Yet even as the blacklisting stands, President Trump suggested last week that Anthropic was 'shaping up' — language that leaves open a path back, contingent on whether the two sides can bridge their disagreement over the terms of military AI use. For now, the Pentagon has moved forward without them, placing its bet on breadth over depth.
The Pentagon announced Friday that it had secured agreements with seven artificial intelligence companies to integrate their systems into the military's most sensitive classified networks. The roster includes SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, and Amazon Web Services—a deliberate effort by the Defense Department to spread its AI capabilities across multiple vendors rather than concentrate them with any single player.
The move represents a significant expansion of the Pentagon's AI footprint. Over the past five months, the department's main platform, GenAI.mil, has been accessed by more than 1.3 million Defense Department personnel. The new agreements will push these seven companies' tools into the Pentagon's Impact Levels 6 and 7 network environments, the classification tiers reserved for the most sensitive military work. Several of these firms already maintain relationships with the Pentagon; Google, for instance, has been operating within Defense Department systems and has now formalized its role in classified operations.
What stands out most sharply in the Pentagon's announcement is what it does not include. Anthropic, the AI company behind Claude, is conspicuously absent from the list. The omission is not accidental. Earlier this year, the Pentagon designated Anthropic a supply-chain risk and barred its use across the Defense Department and its contractor network—a significant move given that Anthropic's tools have been widely deployed throughout military operations. The blacklisting stems from a fundamental disagreement over how the military should be permitted to use Anthropic's AI systems. The company has insisted on guardrails and restrictions; the Pentagon has resisted what it views as constraints on its operational freedom.
The tension deepened with the emergence of Mythos, Anthropic's artificial intelligence model designed with advanced cyber capabilities. The system's power to amplify hacking abilities alarmed both Pentagon officials and corporate security leaders. Emil Michael, the Defense Department's chief technology officer, characterized Mythos as a "separate national security moment" in an interview with CNBC on Friday, even as he acknowledged that Anthropic itself remains classified as a supply-chain risk in the Pentagon's assessment.
Yet the door may not be entirely closed. President Trump remarked last week that Anthropic was "shaping up" in the eyes of his administration, language that signals a potential pathway for the company to reverse its Pentagon blacklisting. Whether that reversal materializes depends on whether Anthropic and the Defense Department can resolve their fundamental disagreement about the terms under which military personnel should access the company's AI tools. For now, the Pentagon has moved forward without it, betting that diversification across seven major players will reduce its dependency on any single vendor while managing the national security risks that advanced AI systems present.
Citas Notables
Anthropic is still a supply chain risk, but Mythos is a separate national security moment— Emil Michael, Defense Department CTO
Anthropic was shaping up in the eyes of his administration— President Trump, on potential reversal of Pentagon blacklisting
La Conversación del Hearth Otra perspectiva de la historia
Why exclude Anthropic when it's already so widely used across the Pentagon?
Because the Pentagon and Anthropic fundamentally disagree about what guardrails should govern military use of AI. Anthropic wants restrictions; the Pentagon wants freedom of operation. That's not a technical problem—it's a values conflict.
And the Mythos model—what makes it different enough to warrant its own security designation?
Mythos was built specifically to excel at cyber operations. It can amplify hacking capabilities in ways that alarmed both the Pentagon and private sector security teams. That's not just another AI tool—it's a force multiplier for adversaries if it falls into the wrong hands.
So why would Trump signal that Anthropic might get back in?
Political leverage, probably. Anthropic is a major player in the AI space. Keeping them permanently blacklisted creates friction with a company that has real influence. If Anthropic bends on the guardrails question, Trump's administration might see an opening to normalize the relationship.
Does spreading AI across seven companies actually reduce risk, or just distribute it?
It does both. You're not dependent on one vendor's security posture or one company's willingness to work with you. But you're also managing seven different systems, seven different integration points, seven different potential vulnerabilities. The Pentagon is betting that diversity outweighs complexity.
What happens to the 1.3 million users already on GenAI.mil?
They get access to more tools, more models, more options. But they also lose access to Anthropic's systems unless and until that blacklisting reverses. For some workflows, that's a real constraint.