Amodei: 'We Don't Know Exactly How' Claude Was Used in Strike That Killed 120 Children
Anthropic CEO Dario Amodei said in a Bloomberg interview Wednesday that Anthropic cannot determine exactly how the Claude AI model was used in a US military strike on an Iranian elementary school that killed approximately 120 children.
“We don’t know exactly how,” Amodei said when asked about Claude’s role in the strike.
The admission is significant: Anthropic has positioned itself as the frontier AI lab most focused on safety and responsible deployment. The disclosure that a Claude model was involved in a lethal military operation where the company itself cannot account for the specifics of that use cuts against the controllability argument that underpins much of Anthropic’s safety research.
What Amodei Said
In the Bloomberg interview, Amodei acknowledged that “terrible” mistakes may be made in military uses of AI. He argued that Anthropic has tried to establish limits and “red lines” around how Claude can be deployed, and that those limits are maintained even if enforcing them risks the company’s commercial future.
That framing positions the red lines as principled rather than performative. The claim is that Anthropic is willing to lose business over them. The evidence against that position is that Claude was in the deployment chain for a strike that Anthropic’s CEO cannot trace after the fact.
The Gap Between Policy and Deployment
AI model providers occupy an unusual position in military use cases. Once a model is licensed and integrated into a government or defense system, the provider typically has no real-time visibility into how it is being used, what decisions it is informing, or what safeguards are applied downstream by the integrating party.
That gap between published policy and operational reality is not unique to Anthropic. The same problem exists across the frontier AI landscape, where multiple labs have signed or are in discussions on government AI contracts. The Iran school strike makes the gap concrete.
Anthropic’s acceptable use policy prohibits using Claude to plan or assist in attacks on civilians. Whether that prohibition was violated, who bears responsibility, and what Anthropic can actually enforce after model deployment are all open questions the Bloomberg interview did not resolve.
Context: Military AI in 2026
The strike disclosure comes during a period of significant expansion in military AI deployment. The NSA is using Anthropic’s Mythos model for cyber work. The Pentagon has been evaluating OpenAI, Google, and xAI models across multiple contract competitions. OpenAI launched GPT-5.5-Cyber for critical infrastructure defenders, and Anthropic launched the Glasswing coalition for AI vulnerability research.
Amodei published a policy manifesto on June 10 calling for mandatory binding regulation of frontier AI, including testing requirements and deployment blocks for high-risk applications. The gap between that call for external constraints and the disclosed military use of Claude without apparent visibility into outcomes illustrates what those constraints would need to close.
No frontier AI lab has yet demonstrated a technical mechanism that lets them trace, let alone prevent, downstream harm once a model is deployed in a third-party system.