US military uses Anthropic's Claude for AI-driven strike planning in Iran war

US Military Leverages Anthropic’s Claude AI for Precision Strike Planning in Simulated Iran Conflict

In a significant development at the intersection of artificial intelligence and military strategy, the United States military has integrated Anthropic’s Claude AI model into high-stakes wargaming exercises. Specifically, during the US Air Force’s “Rampant Lion 2024” simulation, Claude 3.5 Sonnet was employed to formulate detailed airstrike plans targeting Iranian nuclear facilities. This application underscores the accelerating adoption of large language models (LLMs) in defense operations, particularly for complex targeting and resource allocation tasks.

The revelation comes from a detailed report published by the Center for a New American Security (CNAS), a prominent think tank focused on national security and emerging technologies. The exercise, conducted earlier this year, simulated a full-scale aerial campaign against Iran’s hardened nuclear infrastructure, including deeply buried sites like Fordow and Natanz. Planners tasked Claude with generating executable strike packages under tight time constraints, mimicking real-world combat scenarios where rapid decision-making is paramount.

Claude’s performance was noteworthy. In one documented instance, the AI produced a comprehensive 48-hour strike plan within minutes, outlining over 50 targets, assigning specific weapons such as Joint Direct Attack Munitions (JDAMs) and Massive Ordnance Penetrators (MOPs), and sequencing sorties for more than 200 aircraft. The plan incorporated terrain analysis, threat assessments from Iranian air defenses, and collateral damage estimates. According to CNAS analysts, this output surpassed the efforts of junior human staffers, who required hours to achieve comparable results. Senior officers reviewed and refined the AI-generated plans, validating their tactical soundness.

The technical implementation relied on Claude 3.5 Sonnet, Anthropic’s advanced multimodal LLM known for its reasoning capabilities and constitutional AI safeguards. Users interfaced with the model through a custom prompt engineering framework, providing structured inputs such as target coordinates, intelligence summaries, and mission objectives. The AI then leveraged chain-of-thought reasoning to break down the problem: identifying high-value targets, prioritizing based on strategic impact, modeling suppression of enemy air defenses (SEAD), and optimizing fuel and munitions logistics. Outputs were formatted in military-standard briefing slides, complete with maps and timelines, facilitating seamless integration into command-and-control systems.

This exercise highlights several key advantages of AI in military planning. First, speed: Claude compressed what traditionally takes days into minutes, enabling iterative what-if scenarios. Second, consistency: The model applied uniform analytical rigor, reducing human biases like fatigue or overconfidence. Third, scalability: It handled voluminous data—satellite imagery descriptions, order-of-battle intel, and weather forecasts—far beyond individual analyst capacity. In the simulation, Claude proposed innovative tactics, such as staggered low-level ingress routes to evade radar, which human planners later adopted.

However, the integration raises critical technical and ethical considerations. Claude’s safeguards, designed to refuse harmful requests, were tested in this adversarial context. Anthropic’s terms of service prohibit military use, yet the military accessed the model via commercial APIs, prompting debates on dual-use technology governance. CNAS noted instances where Claude flagged ethical risks, such as potential civilian casualties, and suggested mitigations like precision-guided munitions. Reliability remains a concern; while effective in simulation, real-world variables like electronic warfare or degraded communications could degrade performance. The report emphasizes the need for human oversight, positioning AI as an augmentative tool rather than an autonomous decision-maker.

Broader implications extend to force structure and training. The Air Force’s successful trial suggests LLMs could augment the targeting cycle in operations like those against ISIS or Houthi rebels, where persistent ISR (intelligence, surveillance, reconnaissance) feeds demand real-time analysis. Yet, adversaries like China and Russia are pursuing similar capabilities, with reports of their own AI-enhanced command systems. This arms race in AI warfare necessitates robust validation protocols, including red-teaming for hallucinations or adversarial prompts.

Anthropic has not publicly commented on the military’s use, but CEO Dario Amodei has previously discussed AI safety in defense contexts, advocating for aligned systems that prioritize human values. The company’s Claude models incorporate layered safety mechanisms, including refusal training on violent content and scalable oversight techniques. In the wargame, these features prompted Claude to query ambiguities in prompts, such as clarifying rules of engagement, enhancing plan robustness.

Looking ahead, the US Department of Defense’s Chief Digital and Artificial Intelligence Office (CDAO) is scaling such pilots. Initiatives like the Replicator program aim to deploy thousands of AI-enabled drones, where planning tools like Claude could orchestrate swarms. CNAS recommends establishing AI red lines, such as prohibiting lethal autonomous weapons, while investing in explainable AI to build trust.

This episode in “Rampant Lion 2024” marks a milestone in AI’s maturation for warfighting. By demonstrating Claude’s prowess in strike planning against a peer-like threat—Iran’s integrated air defenses—the military has validated generative AI as a force multiplier. As simulations evolve into live exercises, the balance between technological edge and ethical restraint will define the future of AI-driven warfare.

Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.

What are your thoughts on this? I’d love to hear about your own experiences in the comments below.