1 hour ago
19
Dario Amodei, the CEO of Anthropic, said he does not know whether his company’s artificial intelligence model was involved in the missile strikes against Iran.
The admission lands in the middle of an escalating confrontation between one of the world’s most prominent AI safety companies and the US Department of Defense, a conflict that has already resulted in Anthropic being banned from Pentagon systems and labeled a supply-chain risk.
The Pentagon ultimatum and Anthropic’s refusal
The sequence of events that led to this moment started on February 24, 2026, when Defense Secretary Pete Hegseth issued an ultimatum to Anthropic demanding unrestricted access to its Claude AI model. At that point, Claude had already been integrated into US military intelligence workflows used for target identification.
Two days later, on February 26, Anthropic publicly drew its red lines. The company stated it would exclude mass domestic surveillance and fully autonomous weapons systems from its defense contracts. That’s a notable stance from a company that had already been working with the military, suggesting the Pentagon was pushing for capabilities that crossed those boundaries.
The Pentagon’s response was swift and blunt. Officials labeled Anthropic a supply-chain risk and banned its technology from defense systems.
Strikes on Iran and the knowledge gap
In early March 2026, US and Israeli military operations against Iran commenced. The initial onslaught reportedly targeted around 1,000 locations within the first 24 hours. That’s a staggering number of targets to identify, verify, and prioritize in a compressed timeline, exactly the kind of task that AI-assisted intelligence workflows are designed to accelerate.
Claude was already embedded in the military’s targeting infrastructure before the ban took effect. The gap between Anthropic’s refusal and the Pentagon’s subsequent removal of Claude from its systems creates a murky window. Did military operators still have access to Claude’s capabilities during the planning or execution phases of the Iran strikes? Was legacy data processed through Claude models still informing target lists?
Amodei apparently cannot answer these questions with certainty. When the CEO of the company that built the model cannot confirm whether his technology contributed to a military campaign that struck a thousand targets, it reveals something fundamental about how deeply AI has already been woven into national security infrastructure.
Amodei has stated that frontier AI models are inadequate for fully autonomous weapons deployment. That’s a technical claim about current capabilities, not a moral red line. It means the models aren’t good enough yet to pull the trigger on their own, but they can absolutely help humans decide where to aim.
The AI-defense relationship fractures
Anthropic has long positioned itself as the safety-first alternative in the AI arms race. The company was founded by former OpenAI researchers who left over concerns about that organization’s approach to safety. Building an identity around responsible AI development while simultaneously embedding your model in military targeting systems was always going to be a difficult balancing act.
The February confrontation with the Pentagon suggests that balance finally tipped. Hegseth’s demand for unrestricted access and Anthropic’s refusal represent a clean break, but the damage was already done. The model had been inside defense workflows. The targets had been analyzed. And now Amodei is left saying he doesn’t know the full extent of his technology’s role in a shooting war.
Google famously pulled out of Project Maven in 2018 after employee protests over AI-powered drone surveillance. The Anthropic situation is qualitatively different. Google’s Maven involvement was about image recognition for analysis. Anthropic’s Claude was integrated into active targeting workflows during a period that led directly to large-scale military strikes.
What this means for investors and the AI market
The Pentagon’s willingness to label a major AI provider a supply-chain risk and cut it off entirely signals that the government views these companies as replaceable, not indispensable. AI companies that assumed defense contracts would be a reliable, high-margin revenue stream now face the reality that those contracts come with conditions that may be incompatible with their public safety commitments.
With Anthropic banned from Pentagon systems, rival AI firms, particularly those with fewer reservations about unrestricted military use, stand to capture that contract revenue. Companies like Palantir, which have built their entire business model around government and defense applications, are the obvious beneficiaries.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·