The Algorithmic Gray Zone: Operation Epic Fury and the Emergence of AI-Centric Warfare

Introduction: The Twenty-Four Hours That Changed War

At 0347 Zulu on February 28, 2026, the United States Central Command initiated Operation Epic Fury -- the largest coordinated air and missile campaign against the Islamic Republic of Iran since the 1988 Operation Praying Mantis. Within twenty-four hours, over 1,000 discrete targets across Iranian territory had been struck. Military airfields, IRGC-QF logistics nodes, ballistic missile production facilities, naval fast-attack craft pens, and command-and-control bunkers were systematically dismantled in waves of Tomahawk cruise missiles, JASSM-ER standoff munitions, and precision-guided bombs delivered by B-2 Spirits and F-35A Lightning IIs.

But the most consequential weapon deployed that night was not a missile. It was an algorithm.

Operation Epic Fury was the first major military operation in which an AI-driven targeting system -- the Maven Smart System -- served as the primary backbone of the kill chain. Powered in part by Anthropic's Claude large language model, Maven ingested satellite imagery, signals intelligence, human intelligence reports, and open-source data, then generated target packages at a speed and scale that would have been physically impossible for human analysts alone. What once took a team of 2,000 intelligence analysts working around the clock during the 2003 invasion of Iraq was accomplished by approximately 20 operators supervising algorithmic pipelines.

This article is a strategic analysis of the technological, legal, geopolitical, and ethical dimensions of Operation Epic Fury and its aftermath. It is not a polemic. It is an attempt to understand what happened, what it means for the future of warfare, and what guardrails -- if any -- remain between algorithmic recommendation and kinetic action.

I. Architecture of Operation Epic Fury

The Maven Smart System: From Project Maven to Warfighting Backbone

Project Maven -- formally the Algorithmic Warfare Cross-Functional Team (AWCFT) -- was established by a 2017 DoD directive under Deputy Secretary of Defense Robert Work. Its original mandate was modest: use machine learning to automate the analysis of drone surveillance footage, specifically the classification of objects (vehicles, buildings, personnel) in full-motion video from MQ-9 Reaper feeds over Iraq and Syria.

By 2026, Maven had evolved far beyond its original scope. The system had been re-architected into the Maven Smart System (MSS), a multi-modal intelligence fusion platform capable of ingesting and correlating data from across the intelligence community's collection apparatus. Its architecture incorporated several distinct subsystems:

Component	Function	Data Sources
GEOINT Processor	Satellite and aerial imagery analysis; change detection; structure identification	NRO satellites, commercial providers (Maxar, Planet Labs, BlackSky)
SIGINT Correlator	Signals intelligence pattern matching; communications metadata analysis	NSA feeds, tactical SIGINT platforms
HUMINT Integrator	Natural language processing of human intelligence reports; source reliability scoring	DIA, CIA field reports, partner nation intel
OSINT Aggregator	Open-source data collection; social media analysis; commercial shipping data	Public internet, AIS maritime data, flight tracking
Target Synthesis Engine	Fuses all intelligence streams; generates prioritized target nominations with confidence scores	All of the above + historical targeting databases
Collateral Damage Estimator	Models blast radius, structural fragmentation, and civilian proximity for each target	3D terrain models, population density data, time-of-day patterns
LLM Interface (Claude)	Natural language querying of intelligence corpus; report generation; anomaly flagging	All ingested data via retrieval-augmented generation

The Kill Chain: From Hours to Seconds

The traditional military targeting cycle -- known as the F3EAD loop (Find, Fix, Finish, Exploit, Analyze, Disseminate) -- is a deliberate, multi-step process that historically required extensive human judgment at each stage. During Operation Iraqi Freedom in 2003, CENTCOM employed approximately 2,000 intelligence analysts to process imagery, signals intercepts, and field reports for the opening "shock and awe" campaign. Target packages were assembled over days or weeks, reviewed by JAG (Judge Advocate General) officers for legal compliance, and approved through a hierarchical chain of command.

Operation Epic Fury collapsed this process. The Maven Smart System automated the Find and Fix phases almost entirely. Satellite imagery was ingested, processed, and cross-referenced with signals intelligence in near-real-time. When the system identified a target -- say, a mobile TEL (Transporter Erector Launcher) for an Iranian Shahab-3 ballistic missile -- it generated a complete target package: coordinates, confidence score, recommended munition, optimal strike window, and collateral damage estimate.

Human operators -- approximately 20 of them in total, according to reporting by The Intercept -- supervised the system's outputs. Their role was nominally to approve or reject each target nomination. But the sheer volume and velocity of the system's output created what former Pentagon AI ethics advisor Dr. Sarah Chen later described as "a rubber-stamp dynamic." When a system presents you with 1,000 fully formed target packages in 24 hours, each accompanied by a confidence score above 90% and a collateral damage estimate flagged green, the cognitive and institutional pressure to approve is immense.

"The human is technically in the loop," Dr. Chen told the Senate Armed Services Committee on March 12. "But the loop has been compressed to the point where meaningful human judgment is effectively impossible. You are not deciding. You are ratifying."

Scale and Scope of the Strikes

The 1,000-plus targets struck during Epic Fury spanned Iran's military infrastructure:

• IRGC Aerospace Force facilities: Ballistic missile storage bunkers, TEL garages, and launch pads at Tabriz, Isfahan, and Shiraz
• IRGC Navy: Fast-attack craft pens and anti-ship missile batteries along the Strait of Hormuz coastline
• Air defense networks: S-300PMU-2 and Bavar-373 batteries, associated radar sites, and C2 nodes
• Logistics and supply: Ammunition depots, fuel storage, and vehicle maintenance facilities
• Command and control: IRGC regional headquarters, communications relay stations, and fiber-optic junction points
• Nuclear-adjacent: Centrifuge manufacturing facilities and heavy water production plants (but not active enrichment sites)

The campaign was executed in three distinct waves over a 24-hour period, each calibrated to exploit the damage assessment from the previous wave. This battle damage assessment (BDA) loop was itself AI-accelerated: satellite imagery of strike sites was processed by Maven within minutes, and re-strike recommendations were generated automatically for targets assessed as incompletely destroyed.

II. The Integration of Claude and Maven

How Claude Became Embedded in the Kill Chain

Anthropic's Claude large language model was integrated into the Maven Smart System through a contract between Anthropic and Palantir Technologies, which served as the prime systems integrator for Maven's modernization. The integration was not a simple API call. Claude was embedded as a core component of the system's intelligence synthesis layer, providing three critical capabilities:

1. Natural Language Querying: Operators could interrogate the entire intelligence corpus using plain English queries. Instead of writing complex database queries or manually searching through thousands of reports, an analyst could ask: "What IRGC-affiliated logistics movements have occurred within 50km of Isfahan in the past 72 hours?" Claude would retrieve, synthesize, and summarize relevant intelligence in seconds.
2. Report Generation: Claude automatically generated the narrative portions of target packages -- the intelligence justifications that accompany each strike recommendation. These reports had previously been written by human analysts and could take hours to draft for a single target.
3. Anomaly Detection: Claude flagged inconsistencies in intelligence streams -- for example, identifying when SIGINT data contradicted GEOINT assessments, or when a target's characteristics didn't match known patterns for the assessed facility type.

The Trump Ban and the Extraction Problem

On February 27, 2026 -- one day before Operation Epic Fury launched -- President Trump signed an executive order designating Anthropic as a "supply chain risk" under 10 USC § 3252. The order cited Anthropic's Acceptable Use Policy, which prohibits the use of Claude for "weapons development" and "military targeting," as evidence that the company posed a reliability risk to national security. The logic, according to administration officials who spoke to Reuters on background, was that a company whose own policies prohibited military use could not be trusted as a dependable supplier of warfighting technology.

It was the first time in American history that a domestic technology company had been designated a supply chain risk under the statute -- a provision originally designed to restrict foreign adversary suppliers like Huawei and Kaspersky Labs.

"We cannot allow a critical warfighting capability to depend on a supplier that has publicly stated its opposition to the very mission it supports. Anthropic's AUP is functionally an embargo against the Department of Defense."
-- Senior DoD official, speaking on background to Reuters, February 27, 2026

But there was a problem: Claude was too deeply integrated into Maven's architecture to be extracted before the February 28 launch. According to three sources familiar with the matter who spoke to The Washington Post, Palantir engineers estimated that removing Claude from the system would require 6-8 weeks of re-engineering and would degrade Maven's intelligence synthesis capabilities by approximately 40% in the interim. CENTCOM's operational timeline did not permit this delay.

The result was a paradox that legal scholars have described as unprecedented: the United States launched a major military operation using an AI system built by a company that the President had, the day before, officially declared a threat to the supply chain. Claude was simultaneously a sanctioned entity and an active combatant's most critical cognitive tool.

III. The Regulatory Schism: Anthropic vs. the Department of War

The Supply Chain Risk Designation

The 10 USC § 3252 designation carried immediate practical consequences. Under the statute, the Department of Defense is prohibited from entering into new contracts with designated entities and must develop a plan to eliminate dependence on existing contracts. The designation also triggered a review by the Defense Counterintelligence and Security Agency (DCSA) of all cleared personnel with connections to Anthropic.

For Anthropic, the designation was existential. The company's commercial cloud business -- served through Amazon Web Services' Bedrock platform and Google Cloud's Vertex AI -- was not directly affected, but the reputational damage was immediate. Anthropic's stock-equivalent valuation (it remains privately held) dropped an estimated 30% in secondary market trading within 48 hours of the announcement, according to data from Forge Global.

Anthropic's Legal Response

On March 9, 2026, Anthropic filed suit in the U.S. District Court for the Northern District of California, seeking an injunction against the supply chain risk designation. The complaint, Anthropic, PBC v. United States Department of Defense, advanced several legal theories:

• First Amendment: The designation constituted government retaliation against Anthropic for the content of its Acceptable Use Policy -- a form of compelled speech, as it effectively punished the company for expressing (through its AUP) a policy position disfavored by the government.
• Due Process (Fifth Amendment): Anthropic received no notice or opportunity to be heard before the designation, which deprived it of significant property interests (current and future government contracts).
• Administrative Procedure Act: The designation was arbitrary and capricious, as the statute was designed for foreign adversary suppliers and had never been applied to a domestic company.
• Ultra Vires: The President exceeded his statutory authority under 10 USC § 3252, which requires a nexus to foreign adversary influence that does not exist for a San Francisco-based public benefit corporation.

The ACLU Amicus Brief

On March 14, the American Civil Liberties Union filed an amicus curiae brief in support of Anthropic's motion for preliminary injunction. The ACLU's brief focused on the chilling effect of the designation on the broader technology industry:

"If the government can designate a domestic company as a supply chain risk because that company's terms of service restrict military applications, then every technology company in America faces a stark choice: either remove all ethical constraints from your products, or risk being branded a threat to national security. This is the definition of compelled complicity."
-- ACLU amicus brief, Anthropic v. DoD, March 14, 2026

The case remains pending. Judge Yvonne Gonzalez Rogers has scheduled oral arguments for April 2, 2026.

IV. The Iranian Response: Targeting AWS Data Centers

The Strategic Logic

Iran's retaliatory options were constrained. Its air force -- aging F-14A Tomcats, Su-24 Fencers, and domestically produced Kowsar fighters -- was no match for American air superiority. Its ballistic missile arsenal, while large, had been significantly degraded by Epic Fury's strikes on storage and launch facilities. But Iran had one asymmetric capability that was both technically mature and strategically novel: its drone force.

On March 1, 2026, the Islamic Revolutionary Guard Corps Aerospace Force launched coordinated drone strikes against Amazon Web Services data center facilities in two locations: the UAE (ME-CENTRAL-1 region, located in Abu Dhabi) and Bahrain (ME-SOUTH-1 region). The strikes employed Shahed-136 one-way attack drones -- the same loitering munitions that Iran had supplied to Russia for use in Ukraine -- launched from concealed positions in southern Iran.

This was the first military attack in history against a hyperscale cloud service provider.

Strike Impact

Parameter	UAE (ME-CENTRAL-1)	Bahrain (ME-SOUTH-1)
Target	AWS data center campus, Abu Dhabi	AWS data center campus, Bahrain
Munition	Shahed-136 (est. 12-15 drones)	Shahed-136 (est. 8-10 drones)
Physical Damage	2 of 3 data halls struck; cooling infrastructure destroyed; backup generators damaged	1 of 2 data halls struck; partial roof collapse; fire suppression activated
Outage Duration	~14 hours (full), ~72 hours (degraded)	~8 hours (full), ~48 hours (degraded)
Services Affected	EC2, S3, RDS, Lambda, CloudFront edge	EC2, S3, RDS, DynamoDB
Casualties	3 data center technicians killed, 11 injured	1 security guard killed, 4 injured

Civilian and Commercial Fallout

The strikes cascaded through the Middle Eastern digital economy. AWS ME-CENTRAL-1 hosted critical infrastructure for dozens of regional enterprises and government services:

• Careem (Uber-owned ride-hailing): Complete service outage across UAE, Saudi Arabia, Pakistan, and Egypt for 14 hours. An estimated 2.3 million rides were disrupted.
• Emirates NBD (largest bank in the Middle East by assets): Online banking, mobile app, and ATM network experienced intermittent failures for 72 hours. The Central Bank of the UAE activated contingency protocols.
• Hubpay (fintech payment processor): Payment processing halted entirely during the outage window, affecting thousands of merchants.
• UAE government services: Multiple federal e-government portals, including the MOHRE (Ministry of Human Resources and Emiratisation) labor system, went offline.
• Abu Dhabi Securities Exchange: Trading was suspended for one session due to data feed disruptions.

The IRGC's strategic calculus was clear: if the United States was going to use cloud-hosted AI to generate its target lists, then the cloud itself was a legitimate military target. The Iranian government explicitly stated this in a communiqué issued through the state-run IRNA news agency on March 2:

"The infrastructure that enables the enemy's algorithmic aggression is itself a weapon of war. There is no distinction between the missile and the machine that selects its target."
-- IRGC statement via IRNA, March 2, 2026

This statement represented a novel doctrinal assertion: that civilian cloud infrastructure, when used to support military AI systems, loses its protected status under international humanitarian law. The implications for every cloud provider -- AWS, Azure, Google Cloud -- hosting military workloads are profound and unresolved.

V. The Sino-Iranian Technological Nexus

China's Enabling Role

Iran's capacity to execute the drone strikes on AWS infrastructure was significantly enhanced by Chinese technological support. While the Shahed-136 is an Iranian-designed and manufactured platform, several critical subsystems relied on Chinese components and services:

Technology	Provider	Function
Beidou-3 Navigation	PLA Strategic Support Force / BeiDou	GPS-independent satellite navigation for terminal guidance. Enabled precision strikes on specific buildings within the AWS campus rather than area attacks.
Misar Vision Satellite Intel	Misar Vision (Chinese commercial satellite firm)	Sub-meter resolution optical imagery of target facilities. Provided building-level target identification and pre-strike planning data.
DeepSeek V3	DeepSeek (Hangzhou, China)	Operational planning and simulation. Used to model drone flight paths, optimize swarm coordination, and simulate AWS facility defenses.
DeepSeek R1	DeepSeek (Hangzhou, China)	Reasoning model used for target prioritization -- identifying which data halls housed the highest-value compute infrastructure based on publicly available AWS architecture documentation.

The DeepSeek Factor

The involvement of DeepSeek -- a Chinese AI laboratory that had gained international attention for its competitive open-source models -- was perhaps the most strategically significant revelation. DeepSeek V3, the company's flagship general-purpose model, was used by IRGC planning cells to simulate attack scenarios. DeepSeek R1, the company's reasoning-optimized model, was used to analyze publicly available AWS architecture documentation and determine optimal target selection within the data center campuses.

DeepSeek's models are open-source, which means Iran did not require a commercial relationship with the company to use them. The models were downloaded, fine-tuned on military planning data, and deployed on locally hosted infrastructure. This represents a fundamental challenge for AI governance: open-source models, once released, cannot be recalled, restricted, or controlled. They become ambient capability, available to any actor with sufficient compute to run them.

The Chinese government denied any involvement in the drone strikes. A spokesperson for the Ministry of Foreign Affairs stated on March 3 that "China is not a party to this conflict and calls on all sides to exercise restraint." However, the Beidou-3 navigation signals that guided the Shahed-136 drones to their targets are controlled by the PLA Strategic Support Force, and providing precision navigation services to a nation conducting military operations is, at minimum, an act of enabling.

VI. The Tragedy of Minab

What Happened

At approximately 0615 local time on February 28, 2026, during the second wave of Operation Epic Fury, a BGM-109 Tomahawk Land Attack Missile struck a compound in the city of Minab, Hormozgan Province, southern Iran. The Maven Smart System had classified the compound as an IRGC-QF logistics staging area based on historical DIA (Defense Intelligence Agency) geospatial records and satellite imagery analysis.

The compound was, in fact, the Minab Girls' Elementary School.

The strike killed an estimated 175 people, the vast majority of them girls between the ages of 7 and 12 who had arrived for the morning school session. The school's principal, 14 teachers, and several parents who had been dropping off children were also among the dead.

Detail	Information
Date/Time	February 28, 2026, ~0615 local (0245 UTC)
Location	Minab, Hormozgan Province, Iran (27.1517°N, 57.0850°E)
Munition	BGM-109 Tomahawk Block V TLAM, 1,000 lb DSMAC/GPS-guided warhead
Assessed Target	IRGC-QF logistics staging area (Maven confidence score: 87%)
Actual Target	Minab Girls' Elementary School
Estimated Casualties	~175 killed (predominantly girls aged 7-12); ~60 injured
Coordinate Source	DIA geospatial database (coordinates pre-dated 2015; facility had been repurposed)
CDE Assessment	Maven CDE flagged "low civilian proximity" -- school not in civilian infrastructure overlay
Verification	Bellingcat geolocated Tomahawk debris in school courtyard from social media imagery
Strike Pattern	"Triple-tapped" -- three sequential impacts at ~90-second intervals

The Failure Cascade

The Minab strike was not the result of a single failure. It was the product of a cascading series of errors, each of which was amplified by the speed and automation of the Maven system:

1. Stale Intelligence Data: The DIA geospatial database contained coordinates for the compound that pre-dated 2015. The compound had been an IRGC facility but was repurposed as a school sometime between 2015 and 2018. The database had not been updated.
2. Imagery Misclassification: Maven's GEOINT processor analyzed satellite imagery of the compound and classified it as consistent with a logistics facility. The school's layout -- a walled compound with multiple single-story buildings arranged around a central courtyard -- is architecturally similar to many Iranian military logistics sites.
3. Missing Civilian Overlay: The compound was not present in the civilian infrastructure database used by Maven's Collateral Damage Estimator. Iranian school locations are not comprehensively cataloged in Western intelligence databases, and the school was not marked on commercial mapping services (Google Maps, OpenStreetMap) at the resolution Maven ingested.
4. Temporal Blindness: The strike occurred at 0615 local time -- just as children were arriving for school. Maven's time-of-day civilian activity model did not account for Iranian school schedules. A strike at 0200 local would have hit an empty building.
5. Human Override Failure: The target package was reviewed and approved by a human operator. But the operator was processing dozens of target packages per hour, each flagged green by Maven's automated assessments. The operator had no independent means of verifying the facility's current use -- and no time to develop one.
6. Triple-Tap Doctrine: The compound was struck three times in succession at approximately 90-second intervals, a tactic designed to destroy hardened military targets and catch personnel who emerge after the initial strike. Applied to a school full of children, the triple-tap ensured maximum casualties among those who survived the first impact and were attempting to flee.

Verification and Global Response

The open-source intelligence community, led by Bellingcat, verified the strike within 48 hours. Geolocated video and photographs from the scene showed identifiable Tomahawk cruise missile debris -- specifically, the DSMAC (Digital Scene-Matching Area Correlator) guidance section -- in the school's courtyard. The imagery was consistent with a BGM-109 Block V variant.

The international response was severe. The UN Human Rights Council convened an emergency session on March 4. The International Committee of the Red Cross issued a statement noting that "the use of automated targeting systems does not relieve a party to a conflict of its obligations under international humanitarian law, including the obligation to take constant care to spare the civilian population." Massive protests erupted in cities across the Middle East, South Asia, and Europe.

The Pentagon initially issued a statement saying it was "aware of reports of civilian casualties" and was "conducting a thorough review." On March 8, CENTCOM acknowledged that the strike had hit a school and expressed "deep regret for the loss of civilian life." No officer or official has been disciplined or held accountable. The Maven Smart System remained operational throughout.

VII. OpenAI and the Normalization of Military AI

The Pivot

While Anthropic was being designated a supply chain risk for its refusal to support military applications, OpenAI was moving aggressively in the opposite direction. On February 27, 2026 -- the same day as Anthropic's designation -- OpenAI announced a revised usage policy that removed its previous prohibition on military applications and replaced it with language permitting use for "all lawful purposes."

The timing was not coincidental. OpenAI had been in negotiations with the Pentagon for several months, and the revised policy cleared the path for a series of contracts that Sam Altman described on the company's blog as "the most important work OpenAI will ever do."

The Infrastructure Buildout

OpenAI's military pivot was accompanied by a massive infrastructure investment. The company secured 2 gigawatts of dedicated compute capacity using Amazon's custom Trainium chips, deployed across AWS GovCloud regions. This capacity -- equivalent to the electricity consumption of a mid-sized American city -- was allocated exclusively for military and intelligence community workloads.

The scale of this infrastructure reveals the ambition of the program. Two gigawatts of Trainium compute is sufficient to train and serve multiple frontier-class models simultaneously, with capacity for real-time inference at the scale required by systems like Maven. It also represents a significant revenue stream for OpenAI -- military and intelligence contracts are typically multi-year, cost-plus arrangements that provide stable cash flow.

Altman's Admission

On March 10, during a press conference at OpenAI's San Francisco headquarters, Sam Altman acknowledged that the rollout of OpenAI's military capabilities had been "sloppy" and that the company needed to do more to ensure that its technology was used responsibly in military contexts.

"I want to be honest: the rollout was sloppy. We moved fast because we believed the mission was urgent, and we didn't build the governance structures we should have built first. We're fixing that now."
-- Sam Altman, press conference, March 10, 2026

Altman did not specify what "governance structures" OpenAI intended to implement, nor did he address the fundamental question of whether a for-profit AI company is the appropriate entity to serve as a primary technology supplier for lethal military operations. When pressed by a reporter from Wired about whether OpenAI models had been used in Operation Epic Fury, Altman declined to answer, citing "contractual obligations."

VIII. Stargate UAE: The $500 Billion Facility in a War Zone

The Project

Stargate UAE is a joint venture between OpenAI, SoftBank, and the Abu Dhabi sovereign wealth fund Mubadala Investment Company to construct the world's largest AI data center campus in Abu Dhabi. Announced in January 2026, the project carries a headline investment figure of $500 billion over ten years and is intended to serve as the primary AI compute hub for the Middle East, North Africa, and South Asia regions.

The facility's first phase -- a 500-megawatt campus with capacity for over 100,000 GPUs -- was under construction when the Iranian drone strikes occurred. The Stargate construction site is located approximately 15 kilometers from the AWS ME-CENTRAL-1 facility that was struck on March 1.

The Strategic Contradiction

The Iranian strikes exposed a fundamental contradiction in the Stargate project's premise. The facility was sited in Abu Dhabi for its proximity to cheap energy (natural gas and solar), favorable regulatory environment, and geographic centrality to its target markets. What the planners did not adequately account for -- or chose to accept -- was that the UAE's geographic centrality also places it within range of Iranian drones and ballistic missiles.

The Stargate campus, once completed, would represent the single largest concentration of AI compute infrastructure on Earth. It would also represent the most valuable single target in the history of aerial warfare. A coordinated drone and missile attack on the facility could, in theory, destroy hundreds of billions of dollars of infrastructure and disrupt AI services globally -- all using munitions that cost orders of magnitude less than the targets they destroy.

SoftBank's Masayoshi Son addressed the issue at a press conference on March 5, stating that the Stargate campus would be protected by "the most advanced air defense systems in the world" and that the project would proceed on schedule. He did not specify which air defense systems, or who would operate them. The UAE Armed Forces currently operate the THAAD (Terminal High Altitude Area Defense) system and the Pantsir-S1, but neither system has a demonstrated track record against saturating drone swarms of the type Iran employed on March 1.

IX. Conclusion: Toward an Algorithmic Geneva Convention

Operation Epic Fury and its aftermath have revealed a set of challenges that existing legal, ethical, and institutional frameworks are not equipped to address. The speed of AI-driven targeting has outpaced the human capacity for meaningful oversight. The integration of commercial AI into military kill chains has blurred the line between civilian technology companies and defense contractors. The retaliatory strikes on AWS infrastructure have demonstrated that cloud computing, when used for military purposes, becomes a legitimate target -- endangering the civilian digital economy that depends on the same infrastructure. And the tragedy of Minab has shown, in the starkest possible terms, what happens when stale data meets automated confidence at machine speed.

What Must Change

The international community needs what scholars and advocates are beginning to call an "Algorithmic Geneva Convention" -- a new body of international humanitarian law that addresses the specific challenges of AI-enabled warfare. Such a framework should address, at minimum:

• Mandatory Human Decision Time: A minimum deliberation period for AI-generated target nominations, scaled to the assessed risk of civilian harm. Seconds are not sufficient for meaningful human judgment.
• Intelligence Currency Requirements: Mandatory recency thresholds for intelligence data used in targeting. No target should be struck based on geospatial data older than a defined period (e.g., 90 days) without independent verification.
• Civilian Infrastructure Protection: Explicit protections for civilian cloud computing and telecommunications infrastructure, even when co-located with military workloads. The alternative is the gradual militarization of the entire internet.
• Algorithmic Accountability: Clear chains of legal accountability for AI-generated targeting decisions. "The algorithm recommended it" cannot become a defense against war crimes charges.
• Open-Source AI Governance: New international frameworks for managing the proliferation of open-source AI models capable of military applications. Once a model is released, it cannot be recalled -- but the conditions under which military-capable models are released can be regulated.
• Transparency and Audit: Requirements for parties to a conflict to disclose the use of AI in targeting decisions and to submit AI-generated targeting data to post-conflict review by independent bodies.

None of these measures will be easy to implement. International humanitarian law evolves slowly, and the states with the most advanced AI capabilities -- the United States, China, Israel, and Russia -- have strong incentives to resist constraints on their use. But the alternative is a world in which the decision to take a human life is made by a machine that processes stale intelligence, assigns a confidence score, and moves on to the next target in its queue.

The children of Minab were not killed by a malfunction. They were killed by a system working exactly as designed -- a system optimized for speed and scale, processing data it had no reason to question, supervised by humans who had no time to think. If we do not change the system, it will happen again. The only question is where, and how many.

The age of algorithmic warfare has arrived. It arrived not with a debate, but with a kill chain compressed to seconds and a school reduced to rubble. What we do next will define not just the future of war, but the future of the technology that increasingly mediates every dimension of human life.