On February 28, 2026, the United States military used Anthropic’s Claude AI as a decision-support tool during Operation Epic Fury, a joint U.S.-Israeli offensive against Iran that involved nearly 900 strikes in the first 12 hours. According to reporting from the Wall Street Journal and Reuters, U.S. Central Command relied on Claude for intelligence assessment, target identification, and battle simulation — marking the first publicly confirmed significant use of AI-assisted targeting in actual U.S. military combat operations against a nation-state adversary.
The timing could not have been more fraught. The Pentagon’s use of Claude in the Iran strikes occurred within hours of the Trump administration blacklisting Anthropic on February 27, 2026, after CEO Dario Amodei refused demands for unrestricted military access to the company’s AI tools. That same day, OpenAI announced its own Pentagon deal to fill the gap. The result is a collision of military technology, corporate ethics, and executive power that raises questions not just about how wars are fought, but about who gets to decide the boundaries of AI in combat. This article examines how Claude was used in the Iran strikes, the controversy between Anthropic and the Pentagon, the legal and ethical dimensions of AI-assisted targeting, and what this precedent means for future military operations and civilian oversight.
Table of Contents
- How Did the U.S. Use AI-Assisted Targeting Systems in the Iran Strikes?
- The Anthropic-Pentagon Standoff and the Blacklisting
- What Operation Epic Fury Actually Looked Like
- The Legal and Ethical Framework for AI in Combat
- The Precedent Problem and the Accountability Gap
- What Anthropic’s Resistance Actually Meant
- Where AI-Assisted Warfare Goes From Here
- Conclusion
- Frequently Asked Questions
How Did the U.S. Use AI-Assisted Targeting Systems in the Iran Strikes?
CENTCOM deployed Claude across three distinct functions during Operation Epic Fury, none of which involved the AI independently pulling a trigger. First, Claude processed intercepts, satellite imagery, and signals intelligence to generate threat evaluations and situational summaries — essentially digesting enormous volumes of raw data faster than human analysts could manage alone. Second, it was used for target identification: locating, prioritizing, cross-referencing, and confirming high-value targets including leadership compounds, military assets, and strategic sites. Third, CENTCOM used Claude for battle simulation, modeling potential outcomes, rehearsing strike sequences, and predicting risks and collateral damage before pilots and operators executed the actual missions. The pentagon has emphasized that Claude served strictly as a decision-support tool. It did not independently control weapons systems or make lethal decisions without human operators in the loop.
This distinction matters legally and ethically, but it also raises a practical question: when an AI system is processing the intelligence, identifying the targets, and simulating the outcomes, how much of the “decision” is really left to the human pressing the button? The comparison to earlier conflicts is stark. In previous operations, target identification and threat assessment relied on teams of analysts working over days or weeks. Here, the AI compressed that timeline dramatically across nearly 900 strikes in half a day. It is worth noting that this was not the Pentagon’s first use of Claude in a live operation. The military had previously deployed the AI tool during the operation to capture Venezuelan President Nicolás Maduro in January 2026, though the iran strikes represent a far larger scale of engagement and the first use against a nation-state adversary’s military infrastructure.
The Anthropic-Pentagon Standoff and the Blacklisting
The backstory to Claude’s use in the iran strikes involves a direct confrontation between the Trump administration and one of America’s leading AI companies. Secretary of War Pete Hegseth had given Anthropic a deadline to allow unrestricted military use of its AI tools. Dario Amodei publicly resisted, drawing lines against domestic mass surveillance and autonomous weapons development. On February 27, 2026 — one day before the Iran offensive began — the Trump administration blacklisted Anthropic after the company refused to comply. However, the blacklisting did not immediately sever CENTCOM’s access to Claude. The military was still actively using the tool during Operation Epic Fury, which launched within hours of the ban order.
This gap between policy announcement and operational reality is not unusual in government — procurement and access systems do not flip off like a light switch — but it creates an awkward situation. The administration was publicly punishing a company whose product it was simultaneously relying on in a major combat operation. If the tool was effective enough to use during the most consequential military action of the year, the ban looks more like political retaliation than a genuine security or capability concern. The immediate beneficiary was OpenAI, which announced a Pentagon deal the same day Anthropic was blacklisted. This raises a different set of concerns. OpenAI has been more willing to work with military and government clients without the guardrails Anthropic insisted on. Whether that flexibility translates to better or worse outcomes for military operations — and for civilians in strike zones — remains to be seen.
What Operation Epic Fury Actually Looked Like
The joint U.S.-Israeli offensive, codenamed Operation Epic Fury by the Americans and Operation Roaring Lion by the Israelis, was massive in scale. Nearly 900 strikes hit Iranian targets in the first 12 hours. The U.S. deployed B-2 stealth bombers, F/A-18 and F-35 fighter jets, suicide drones, and AI tools. B-2 bombers flew directly from the United States to strike hardened underground Iranian missile facilities using 2,000-pound bombs — a mission profile that underscores the severity of the targets and the level of planning involved. The scale of the operation helps explain why AI-assisted targeting was attractive to military planners.
Coordinating 900 strikes across multiple weapons platforms, in concert with a foreign military, against a nation-state with sophisticated air defenses, is an enormously complex logistical and intelligence challenge. Traditional methods of target development and deconfliction would likely have required a longer timeline. Claude’s ability to process signals intelligence, cross-reference target data, and simulate strike outcomes allowed CENTCOM to compress the planning and execution cycle in ways that would have been difficult with human analysts alone. That said, speed is not always a virtue in warfare. One of the longstanding concerns about AI in military targeting is that compressing decision timelines can reduce the opportunity for human judgment to catch errors — misidentified targets, faulty intelligence, or underestimated civilian presence. The Pentagon has not released detailed information about collateral damage assessments from Operation Epic Fury, and independent verification of strike accuracy remains limited.
The Legal and Ethical Framework for AI in Combat
International humanitarian law requires that military strikes distinguish between combatants and civilians, that the expected military advantage be proportional to anticipated civilian harm, and that parties to a conflict take precautions to minimize civilian casualties. None of these obligations disappear because an AI system is involved in the targeting process. The question is whether AI makes compliance easier or harder — and the honest answer is that it depends entirely on how the tool is used. Proponents argue that AI-assisted targeting can improve compliance by processing more data, identifying patterns humans might miss, and modeling collateral damage more accurately than rushed human analysis. Critics counter that AI systems can encode biases present in their training data, that their outputs can create a false sense of precision, and that the speed they enable can outpace meaningful human review.
The comparison to drone warfare is instructive: when the U.S. began widespread drone strikes in the 2000s and 2010s, proponents similarly argued that precision-guided munitions would reduce civilian casualties. Investigations by journalists, human rights organizations, and the military itself later revealed significant undercounting of civilian deaths. The tradeoff is real. If Claude’s battle simulations accurately predicted collateral damage and CENTCOM used that information to adjust strike parameters, the AI may have prevented civilian deaths that would have occurred under a faster, less data-driven approach. If, on the other hand, the AI’s assessments gave planners false confidence in the precision of strikes, the tool may have enabled more aggressive targeting than would otherwise have been approved.
The Precedent Problem and the Accountability Gap
The use of AI-assisted targeting in the Iran strikes sets a precedent that will be difficult to walk back. Once a military demonstrates that AI tools can be integrated into large-scale combat operations, the institutional incentive is to expand their use, not constrain it. Other nations — including China, Russia, and Israel — are developing their own AI military capabilities, and the U.S. use of Claude in a live combat operation will accelerate that global competition. The accountability question is particularly thorny.
When a strike goes wrong and kills civilians or hits the wrong target, the existing military justice and oversight system assigns responsibility to human commanders and operators. But when an AI system identified the target, assessed the intelligence, and simulated the outcome, the chain of responsibility becomes murkier. The Pentagon’s insistence that Claude was a “decision-support tool” with humans in the loop is partly a legal strategy — it preserves the fiction of clear human accountability even as the practical locus of analysis shifts to the machine. There is also a domestic oversight gap. Congressional authorization for the Iran strikes has been debated, and the role of AI in targeting adds another layer that existing oversight mechanisms are not well-equipped to scrutinize. Members of Congress who review strike decisions and intelligence assessments may not have the technical expertise to evaluate whether an AI system’s target recommendations were sound, and the classified nature of the tools makes independent review nearly impossible.
What Anthropic’s Resistance Actually Meant
Anthropic’s refusal to grant the Pentagon unrestricted access to Claude was not a blanket refusal to work with the military. The company had already allowed Claude’s use in the Venezuela operation and, clearly, in preparatory work for the Iran strikes. Amodei drew specific lines: no domestic mass surveillance, no autonomous weapons development.
The distinction matters because it suggests a framework in which AI companies can work with defense clients while maintaining ethical boundaries — at least in theory. The Trump administration’s response — blacklisting the company and immediately turning to a more compliant competitor — sends a clear signal to the rest of the AI industry. Companies that attempt to set conditions on military use of their products risk losing not just defense contracts but broader government business. Whether other AI companies will follow Anthropic’s example or OpenAI’s is an open question, but the financial and political incentives heavily favor compliance.
Where AI-Assisted Warfare Goes From Here
The Iran strikes are a beginning, not an endpoint. The Pentagon’s rapid pivot from Anthropic to OpenAI suggests that the military views AI-assisted targeting as essential to future operations, not experimental. The next frontier is likely greater autonomy — systems that not only identify targets and simulate outcomes but execute decisions with less human involvement.
The Department of Defense’s own AI strategy documents envision a future in which AI is embedded at every level of military planning and execution. For the public, the challenge is maintaining meaningful oversight of a technology that is advancing faster than the legal, ethical, and political frameworks designed to govern it. The use of Claude in Operation Epic Fury happened with minimal public debate, no specific congressional authorization for AI-assisted targeting, and a corporate ethics dispute that was resolved by punishing the company that tried to impose limits. That pattern — technology outrunning accountability — is the real story, and it will only intensify as AI capabilities grow.
Conclusion
The U.S. military’s use of Anthropic’s Claude AI during Operation Epic Fury represents a genuine inflection point. For the first time, AI-assisted targeting was used at scale in combat operations against a nation-state adversary, compressing intelligence assessment, target identification, and battle simulation into a timeline that human analysts alone could not have matched across nearly 900 strikes.
The tool remained under human control, but its influence on the decisions that shaped those strikes was substantial. The surrounding controversy — Anthropic’s resistance, the administration’s blacklisting, OpenAI’s rapid entry — reveals the broader stakes. The question is no longer whether AI will be used in warfare but under what conditions, with what safeguards, and subject to whose oversight. The precedent set in the Iran strikes will shape those answers for years to come, and the early signs suggest that speed, capability, and political compliance are winning out over caution, accountability, and ethical constraint.
Frequently Asked Questions
Did AI independently launch weapons or make kill decisions during the Iran strikes?
No. According to Pentagon statements and reporting from the Wall Street Journal and Reuters, Claude served as a decision-support tool only. It processed intelligence, identified targets, and simulated outcomes, but human operators made the final decisions and controlled weapons systems.
Was the U.S. military still using Claude after the Trump administration blacklisted Anthropic?
Yes. The blacklisting was announced on February 27, 2026, and the Iran strikes began on February 28. CENTCOM was still actively using Claude tools during the operation, within hours of the ban order. Operational access was not immediately severed.
Has the U.S. military used AI-assisted targeting before the Iran strikes?
The Pentagon had previously deployed Claude during the operation to capture Venezuelan President Nicolás Maduro in January 2026. However, the Iran strikes represent the first publicly confirmed use of AI-assisted targeting in combat operations against a nation-state adversary at this scale.
What did Anthropic refuse to do that led to the blacklisting?
Anthropic CEO Dario Amodei refused Pentagon demands for unrestricted military access to Claude. Specifically, he drew lines against allowing the AI for domestic mass surveillance and autonomous weapons development. The company had cooperated with more limited military uses.
Who replaced Anthropic after the blacklisting?
OpenAI announced a Pentagon deal on the same day Anthropic was blacklisted, February 27, 2026, positioning itself to fill the gap in military AI services.
Is AI-assisted military targeting legal under international law?
International humanitarian law does not specifically prohibit AI-assisted targeting, but it requires that all military operations — regardless of the tools used — comply with principles of distinction, proportionality, and precaution. The legality depends on how the AI is used and whether human operators maintain meaningful control over targeting decisions.