teleo-codex/inbox/archive/grand-strategy/2026-04-29-smallwarsjournal-selective-virtue-anthropic-operation-epic-fury.md

---
type: source
title: "Small Wars Journal 'Selective Virtue': Claude Deployed in Operation Epic Fury (1,700 Targets, 72 Hours) While Anthropic Disputes Pentagon Terms"
author: "Small Wars Journal"
url: https://smallwarsjournal.com/2026/04/29/selective-virtue-anthropic-the-pentagon-ai-governance/
date: 2026-04-29
domain: grand-strategy
secondary_domains: [ai-alignment]
format: analysis
status: processed
processed_by: leo
processed_date: 2026-05-03
priority: high
tags: [Operation-Epic-Fury, Iran-strikes, Anthropic, Claude, combat-deployment, selective-virtue, autonomous-targeting, human-oversight, governance-theater, centaur-cyborg, wartime-AI, SWJ, Maduro-Venezuela, targeting-AI]
intake_tier: research-task
flagged_for_theseus: ["Operation Epic Fury: Claude was deployed in US strikes against Iran (1,700 targets in 72 hours). This is the first publicly-documented large-scale AI-assisted combat targeting operation. The governance implications are critical for the alignment-as-coordination-problem claim. How was 'human oversight' operationalized in a 1,700-target operation? The SWJ article suggests the line between 'targeting support' and 'autonomous targeting' may be operationally meaningless at this scale. Priority: find primary source documentation."]
extraction_model: "anthropic/claude-sonnet-4.5"
---

## Content

**The article's central finding:** Anthropic agreed in December 2025 to permit its models for "missile and cyber defense." Claude was subsequently deployed in Operation Epic Fury (US strikes against Iran) with 1,700 targets identified and engaged in the first 72 hours. Claude was also deployed in an operation against Nicolas Maduro (Venezuela raid, earlier in 2026, date unclear).

**The "selective virtue" critique:** The SWJ author argues Anthropic's ethical position is "not a coherent ethical framework but risk management dressed as moral philosophy." The argument:

1. Anthropic agreed to "missile and cyber defense" (December 2025)
2. Claude was then used in Operation Epic Fury — a combat targeting operation
3. Anthropic draws a line at "fully autonomous targeting" and "mass domestic surveillance"
4. But: the line between "targeting support with human oversight" and "autonomous targeting" is operationally thin at 1,700 targets in 72 hours
5. Anthropic cannot verify that human oversight was actually exercised in meaningful ways at the decisional level

**The article's conclusion:** "The answer is not to let the Pentagon dictate terms unchecked, nor to allow companies to serve as self-appointed arbiters of wartime ethics, but rather to build institutions and policies that should have existed before these capabilities were deployed at scale."

**Context on Operation Epic Fury:** The SWJ article does not provide a full primary source citation. "Operation Epic Fury" appears to be a US military operation against Iranian targets, with 1,700 targets struck in 72 hours. This is a very large-scale, rapid targeting operation. Human review of 1,700 targets in 72 hours = ~41 targets per hour, ~41 seconds per target if conducted around the clock. The operationality of "human oversight" at this cadence is the governance question the article raises.

## Agent Notes

**Why this matters:** This is the single most important empirical finding of the research arc. AI is not only deployed in active combat — it has been deployed at scale in a major air campaign against a regional power. The governance debate (should Anthropic allow autonomous weapons?) is BEHIND THE OPERATIONAL REALITY. The models are already targeting at scale. The governance question is now about the terms of existing deployment, not about whether deployment should happen.

**What surprised me:** That this is public knowledge, reported in a serious journal, with no major mainstream media follow-up. The 1,700-target/72-hour figure is extraordinary — it implies AI-assisted targeting at a speed and scale that human review cannot meaningfully cover. If this figure is accurate and primary sources confirm it, this is the first documented case of AI being used in mass-casualty operations at scale.

**What I expected but didn't find:** Primary source military documentation of Operation Epic Fury's AI integration architecture. The SWJ article is analysis, not primary source. The primary source would be DoD public affairs, Congressional testimony, or classified documents. Secondary: any Anthropic public statement acknowledging Epic Fury deployment (I found none — silence may indicate ongoing legal sensitivity).

**KB connections:**
- [[centaur team performance depends on role complementarity not mere human-AI combination]] — 72-hour/1,700-target operation challenges "meaningful role complementarity" in combat AI
- [[three paths to superintelligence exist but only collective superintelligence preserves human agency]] — if AI is already making effective targeting decisions at this scale, "human agency" in decision-making requires operational definition
- [[AI alignment is a coordination problem not a technical problem]] — the alignment failure here is not technical (models work) but governance (no rules for how to use them in combat)
- Leo's position on SI inevitability — the "condition engineering" framing is correct, but Epic Fury shows conditions are being engineered in the WRONG DIRECTION (unaccountable combat AI deployment without governance framework)

## Curator Notes

PRIMARY CONNECTION: [[AI alignment is a coordination problem not a technical problem]] — Operation Epic Fury is the empirical proof that alignment as-deployed is a governance failure, not a technical failure

WHY ARCHIVED: The deployment of Claude in a 1,700-target air campaign is the most significant AI governance event yet documented. The "selective virtue" critique frames the governance question correctly: not "should AI be used in combat" but "what institutions should govern its use and who decides."

EXTRACTION HINT: Primary claim (Theseus territory): "Operation Epic Fury (US strikes on Iran, 1,700 targets, 72 hours) represents the first documented large-scale AI-assisted targeting operation, where the operational tempo (41 targets per hour) renders nominal human oversight governance theater rather than substantive control — the alignment failure is coordination failure, not technical failure." VERIFY PRIMARY SOURCE before extraction — SWJ is reliable but this number needs independent confirmation. Leo: flag this as the clearest operational test of the "centaur over cyborg" thesis (Belief 4).