The Middle East, already a region of instability, has erupted into renewed conflict, potentially marking the first full-scale war of the AI era. Recently, the United States and Israel launched a joint military operation against Iran, codenamed "Epic Fury." In retaliation, Iran launched missile and drone strikes against nine neighboring countries hosting US military bases.
Around the same time, the U.S. Department of War signed a cooperation agreement with AI leader OpenAI to deploy artificial intelligence for military purposes. While these two events may be coincidental in timing, their convergence could hold historical significance. OpenAI accepted the project after its direct competitor, Anthropic, rejected a final ultimatum from the U.S. Department of War. Anthropic refused to grant the U.S. military broad AI access, specifically objecting to its use for domestic surveillance of American citizens and in autonomous weapon systems, leading to its blacklisting by the U.S. government.
Recently, media outlets exposed a 1,600-word internal memo from Anthropic, providing a comprehensive look at this landmark standoff between an AI company and the U.S. government. The memo reveals political logic behind the conflict that is far more complex than mere technical disagreements. In the memo, Anthropic CEO Dario Amodei suggested that his company's poor relationship with the government stemmed from not donating to the White House or offering flattery like some competitors. He accused rivals of performative safety measures, claiming the security precautions in OpenAI's agreement with the Department of War were merely a charade.
**The Pentagon's Designated AI Provider**
Why was Anthropic the chosen provider? Despite not being the largest player in the AI industry, with a significantly smaller consumer user base than OpenAI or Alphabet, Anthropic is the undeniable leader in the enterprise market. In the enterprise LLM sector, Anthropic holds a 32% market share, ranking first. In the highly valuable niche of programming and intelligent workflows, its model Claude commands a dominant 70% market share.
Anthropic's annualized revenue has reached $19 billion, with 80% originating from enterprise clients. Last month, the company completed a $30 billion Series G funding round led by GIC, achieving a valuation of $380 billion. Investors widely believe its "moat" in governance compliance and institutional trust is difficult for competitors like Alphabet and OpenAI to overcome.
Although OpenAI, Alphabet, and xAI all received $200 million contracts from the U.S. Department of War, Anthropic's Claude was the only major model authorized for deployment on the U.S. military's highest-level classified networks. Together with data analytics provider Palantir Technologies Inc. and Amazon.com's AWS cloud service, they form a crucial AI triad for the U.S. military.
Specifically, Palantir provides the big data analytics framework, while Claude acts as its "cognitive engine." This combination allows the military to process vast amounts of satellite imagery and signals intelligence at "machine speed" for rapid combat decision-making. A military operation earlier this year to capture the Venezuelan president served as an example of this powerful partnership. Post-operation disclosures revealed the military used Claude's logical reasoning capabilities on Palantir's platform to quickly analyze intercepted intelligence and satellite images, cross-referencing travel records and financial data to pinpoint the president's real-time location. The operation resulted in dozens of Cuban and Venezuelan soldier casualties, with no U.S. fatalities.
The recent large-scale U.S. and Israeli attack on Iran also utilized the Claude and Palantir combination for intelligence assessment, target identification, and battlefield scenario simulation. In other words, the U.S. government reportedly banned Claude on a Friday, yet used it for the attack on Iran the following Saturday. This creates a significant problem for the Pentagon. Until recently, Claude remained the only AI tool permitted on classified military systems. Government officials acknowledged that Claude outperforms alternatives in certain areas and admitted that removing Anthropic's tools from existing systems would be a highly disruptive process.
**Two AI Usage Red Lines**
Given Claude's effective use in U.S. military operations abroad, why did the company clash with the Department of War? The conflict centers on two fundamental AI usage red lines.
During negotiations, Anthropic insisted on including two explicit prohibitions in any contract: first, a ban on using its AI for large-scale domestic surveillance of U.S. citizens, and second, a ban on its use in fully autonomous lethal weapon systems. The Department of War refused to codify these prohibitions in the contract, maintaining that AI should be available for "any lawful purpose."
Amodei's memo detailed that the Department of War made a final demand: to remove restrictive language concerning "bulk data acquisition for analysis" from the contract. This clause directly addressed Anthropic's primary concern—AI being trained on mass-collected U.S. citizen communication data to enable domestic mass surveillance. Anthropic refused to concede on this point.
The stalemate led the U.S. Department of War to declare Anthropic a "national security supply chain risk" after a deadline passed. A former U.S. President publicly labeled Anthropic a "radical left-wing, politically correct company." Mere hours later, OpenAI announced it had signed the contract that Anthropic rejected.
Industry groups supporting Anthropic have mobilized. A consortium including investors like NVIDIA sent a letter to the Department of War expressing concern over labeling a U.S. company a supply chain risk, a designation typically reserved for firms linked to foreign adversaries. Anthropic has stated it will challenge this designation in court.
Considering 80% of Anthropic's revenue comes from enterprise clients, being placed on a U.S. government "supply chain risk" list could inflict significant financial damage. It also poses a major risk to the company's planned initial public offering this year, as many corporate clients holding government contracts may be forced to abandon Anthropic's products.
**Public Praise, Swift Takeover**
Despite OpenAI CEO Sam Altman publicly supporting Anthropic's stance on safety red lines and praising their commitment on the very day of the announcement, his company swiftly took over the contract Anthropic abandoned. This seamless transition and what appeared to be insincere remarks clearly angered Amodei.
It is noteworthy that Anthropic and OpenAI share a history and rivalry. Anthropic's founding team primarily came from OpenAI, leaving due to disagreements with Altman's product and commercial direction. When OpenAI announced advertising plans, Anthropic spent millions on a Super Bowl ad mocking its rival, prompting an angry response from Altman.
Interestingly, at last month's AI summit in India, Altman and Amodei were placed next to each other by organizers but reportedly ignored one another, refusing even a perfunctory handshake despite a request from the Indian Prime Minister, creating an awkward atmosphere.
The Department of War contract incident has intensified Amodei's anger. In the internal memo, he wrote that OpenAI's deal is mere "security theater," and that Altman "presents himself as a peacemaker and dealmaker," which is a "complete lie." The term "security theater," coined by security expert Bruce Schneier, describes measures that create a feeling of security without providing substantial improvement.
Amodei argued that OpenAI's contract includes seemingly responsible "red lines" against domestic surveillance and autonomous weapons, but these clauses rely entirely on the government's own interpretation of "lawful," with no independent verification or enforcement mechanism, unlike Anthropic's firm stance.
Amodei expressed extreme dissatisfaction, stating bluntly that the real reasons for being targeted were: not donating to a political campaign (while OpenAI executives donated significantly), not offering praise to the administration, supporting AI regulation (which contradicts the administration's agenda), and genuinely upholding AI red lines instead of creating "security theater" to placate employees. He referenced Altman's personal $1 million donation and OpenAI president's subsequent $25 million donation to a political action committee, alongside Altman's public praise for the administration and accompanying the White House on foreign trips. While many tech giants curried favor with the administration through donations, Anthropic maintained its distance, a point Amodei emphasized in his criticism.
**Explanation Faces Widespread Skepticism**
OpenAI, of course, disputes these accusations. The company published a lengthy statement on its website claiming its contract with the Department of War includes three "red lines": not to be used for mass domestic surveillance, not to command autonomous weapons where human control is required, and not to be used for high-stakes autonomous decision-making without human approval.
Altman acknowledged this week that the initial contract had flaws and announced the company had renegotiated terms to include stronger safeguards. He specifically emphasized that contract language was amended to explicitly exclude U.S. domestic surveillance.
However, Anthropic countered, pointing out a fundamental flaw: OpenAI's contract still permits the government to use AI for "all lawful purposes," with the power to define "lawful" resting solely with the U.S. government.
Anthropic is not alone in its skepticism. Many industry analysts doubt OpenAI will uphold its red lines. They argue that if the Department of War had truly accepted these three lines, it would not have clashed with Anthropic in the first place.
A former head of OpenAI's security team questioned on social media why the public should believe that intelligence agencies are excluded from the contract, noting a lack of strong evidence. Critics also pointed out that U.S. intelligence agencies previously justified mass surveillance of millions of ordinary Americans' communications as "lawful" under the vague framework of the Foreign Intelligence Surveillance Act. Anthropic did not abandon its standards for such "lawful" uses, whereas OpenAI blurred this boundary.
A deeper issue concerns the definition of "human oversight." A former member of OpenAI's geopolitics team noted that the defense industry lacks a consensus definition for "meaningful human control" in autonomous weapons. Does a soldier clicking "confirm" on an AI-generated target constitute oversight? Or must each strike decision be individually reviewed? Under the time pressures of modern battlefields, the answer could determine thousands of lives.
Notably, 96 OpenAI employees signed an open letter before the contract was announced, urging management to "continue to refuse the Department of War's current requests to allow the use of our models for domestic mass surveillance and autonomous killing without human supervision."
**Blurred Lines of AI Warfare**
The primary applications of AI in modern warfare are clear: rapid intelligence summarization, target generation, battlefield awareness, and logistical and command decision support. Individually, these functions seem like "assistive tools," but when linked into a complete chain from sensing to lethal action, their nature fundamentally changes. This is the power of the Claude and Palantir combination.
The chain—from sensor data to AI interpretation, target selection, and weapon activation—carries a key risk: the entire process could occur with minimal human control or awareness.
This is not hypothetical. The Lavender system, widely used by Israel in the Gaza conflict, has been described by former intelligence officials as an AI machine that automatically generates hundreds of potential targets daily. Human reviewers reportedly spent an average of less than a minute, sometimes under 20 seconds, reviewing each target before it entered the strike pipeline. In effect, human oversight became a rubber stamp rather than a substantive check.
Beyond autonomous weapons, the controversy raises fundamental questions. Who has the authority to define "lawful"? If AI usage boundaries are self-interpreted by the government, safety guardrails become political tools rather than genuine protections. Anthropic's insistence aimed to codify a written boundary, however imperfect. OpenAI's contractual approach blurs this red line.
Do tech companies have the capability and courage to resist pressure from the world's most powerful government? Each time, companies navigate between public opinion, commercial interests, and government pressure, but their hesitation appears to be diminishing.
In 2018, Google's Project Maven sparked protests from 4,000 employees, leading the company not to renew its contract, which was subsequently taken up by Microsoft and Amazon.com. In 2019, Microsoft employees protested providing HoloLens headsets for battlefield use. In 2024, Google employees protested the Project Nimbus contract with Israel, to no effect.
Last summer, Google, Anthropic, OpenAI, and xAI each received Department of War contracts worth up to $200 million to develop "agentic AI workflows" for combat, intelligence, and enterprise systems, with Anthropic receiving the most sensitive security clearances.
This time, however, the near-simultaneous timing of OpenAI signing the contract and the U.S.-Israel air strikes on Tehran created an unprecedented impact. Regardless of the final outcome, Anthropic's resistance demonstrated an alternative possibility: at least for a moment, one company chose not to compromise. But if one company refuses, there is always a competitor willing to step in.
Comments