NextFin News - Anthropic Chief Executive Dario Amodei said on June 10 that he does not know whether Claude was used in a reported strike on a school in Iran. The answer leaves Anthropic unable to say whether its best-known product was involved in a specific wartime attack.
The issue grew out of reporting and official reaction that began months earlier. The Guardian reported in March that the U.S. military had used Claude during the Iran campaign despite Donald Trump’s order to sever ties with Anthropic, citing reporting by the Wall Street Journal and Axios that said military command used the tools for intelligence work, target selection and battlefield simulations. The Washington Post later reported that, in the first 24 hours of the attack on Iran, the U.S. military used what it described as the most advanced artificial intelligence it had ever deployed in warfare to help strike 1,000 targets.
Those reports did not show that Claude was used in every specific strike. They did establish that Anthropic’s model was already being used within a major military campaign.
That leaves Anthropic with a practical oversight problem, not just a public-relations one. The company can set terms of use, publish safety evaluations and block obvious misuse through its own interface. But when a model is accessed through a government or contractor environment, Anthropic may have far less visibility into how it is used. The reported school strike in Iran, if confirmed, would fall into that gap: a downstream use outside Anthropic’s direct customer interface, but still linked to its brand and underlying technology.
Amodei co-founded Anthropic after leaving OpenAI and has spent years presenting the company as a more cautious developer of frontier AI than some Silicon Valley rivals. Anthropic has promoted Claude as a model with stronger guardrails, particularly around harmful requests, weapons and surveillance. The company’s message has been that powerful AI should come with tighter deployment rules, not looser ones. Against that record, being unable to say whether Claude was involved in a strike on a school raises questions about the safety-first case Anthropic has made to enterprise customers and policymakers.
The available evidence still does not support treating the reported school strike as a settled fact about Claude’s role. The March reporting addressed broader military use in the Iran campaign, not a verified audit trail for one strike. Claude may have been used somewhere in the targeting workflow without being used to assess the specific school in question. Model output may also have been one input among many rather than the deciding factor.
The dispute became more charged in March, when Trump ordered federal agencies to stop using Claude and denounced Anthropic on Truth Social as a “Radical Left AI company.” The Guardian said defense secretary Pete Hegseth later accused Anthropic of “arrogance and betrayal” and demanded full access to the company’s models for lawful purposes. The clash pits a White House and Pentagon that appear increasingly comfortable using AI in military operations against an AI vendor trying to maintain that it does not want its models used for violent ends, weapons development or surveillance.
The business stakes are substantial as well. Anthropic competes in a market where government contracts, enterprise deployments and trust from regulated industries matter alongside model performance. A company seen as essential to national-security customers can gain revenue and standing. A company seen as unable to govern where its tools go, or unable to explain that clearly, can draw scrutiny from customers, lawmakers and employees at the same time. That pressure is sharper for Anthropic because Claude is marketed not only as a capable general-purpose model but as a comparatively responsible one.
The episode also points to a broader problem for AI companies. As frontier models become more capable, the debate is shifting from whether a company supports a use case to whether it can observe and audit that use after deployment. In cloud software, vendors can often log access, revoke permissions and track usage patterns with precision. In military and intelligence settings, those controls are weaker, and access may run through agencies, contractors and secure systems that the original developer cannot see into. Once a model is integrated there, a vendor’s denial or uncertainty does not necessarily mean it played no role; it may mean the vendor no longer has visibility.
For now, the known facts are limited. March reporting tied Claude to the Iran campaign. The U.S. military’s use of frontier AI in that campaign was publicly acknowledged in broad terms. Trump and Hegseth turned the matter into a political fight. And on June 10, Amodei said he did not know whether Claude was used in the reported school strike.
Explore more exclusive insights at nextfin.ai.
