AI Firms' Ethical Red Lines Collapse in Wartime

Opening

Dear reader, you’ve probably seen last week’s news about the US airstrikes on Iran. But something else caught my eye as I read the coverage. The core tool behind an operation that struck more than 1,000 targets in its first 24 hours turned out to be the very same AI model we use every day to draft emails and write code. Anthropic’s Claude, running through Palantir’s Maven Smart System¹, was used by the US military for intelligence analysis, target identification, and combat scenario simulation. And this happened the day after President Trump had declared a ban on Anthropic’s use across federal agencies.

Do the “ethical principles” AI companies have championed actually hold up in wartime? What this episode revealed is a massive gap between principle and reality.

From Project Maven to the Iran War — An Eight-Year Trajectory

This story starts in 2017. The US Department of Defense’s Project Maven was an effort to use AI to identify objects in drone footage, and Google won the contract at the time.

But in March 2018, more than 4,000 Google employees signed an internal petition arguing that “Google’s technology should not be used to kill people,” and a dozen or so employees actually quit over it. Google eventually told the Pentagon it would not continue, and that same year it withdrew from bidding on a $10 billion Pentagon cloud contract, saying it couldn’t be certain the deal aligned with its own AI principles.

Palantir stepped into the space Google vacated. It formed a consortium with AWS to handle software development, but progress was slow. The turning point came in 2024, when Anthropic partnered with Palantir and the Maven Smart System was completed — Claude was placed on top of the Palantir platform.

One thing worth noting here: by the time Anthropic signed that 2024 partnership, it already knew its technology would be integrated into the Defense Department’s drone operations system — the exact role Google had refused eight years earlier.

Anthropic’s Two Red Lines, and the Irony Within Them

In July 2025, Anthropic signed a $200 million contract with the Defense Department. Then in early 2026, Anthropic CEO Dario Amodei presented the Pentagon with two conditions.

First, no mass surveillance of American citizens. Second, no development of fully autonomous weapons — meaning no systems where AI selects and attacks targets without human oversight.

The Pentagon’s response was blunt: for any lawful purpose, that’s our call to make — why should a software vendor have a say in the military’s operational decisions? Deputy Defense Secretary Emil Michael compared it to “a hammer complaining to the carpenter about how it should be swung.”

But look closely at Amodei’s two conditions and an irony emerges. Banning surveillance only of “American citizens” implicitly permits surveillance of foreign nationals. Opposing only “fully” autonomous weapons implicitly allows partially autonomous ones. And in interviews, Amodei explained that he wasn’t opposed to fully autonomous weapons on principle — rather, “the technology isn’t reliable enough yet.” This was less an ethical red line than a technical reservation.

Big Tech on the Eve of War — Four Companies, Four Different Colors

On February 27, when talks between Anthropic and the Pentagon broke down, the situation changed rapidly.

President Trump posted on Truth Social ordering all federal agencies to stop using Anthropic. Defense Secretary Pete Hegseth designated Anthropic a ‘supply-chain risk’ — a label normally reserved for adversary-nation companies like Huawei, and one that had never before been applied to an American company.

That very night, OpenAI announced a deal to deploy in classified environments for the Pentagon. CEO Sam Altman said OpenAI “shares the same red lines as Anthropic,” but the actual contract was structured to permit “all lawful uses.” Altman later admitted the deal “looked opportunistic and sloppy.” There was pushback inside OpenAI too — the head of its robotics division resigned, citing concerns over autonomous weapons and surveillance.

Meanwhile, Google had already quietly removed the weapons-and-surveillance ban from its AI principles back in February 2025 — the very same principles that 4,000 employees had petitioned over in 2018. In its official statement, Google cited “an increasingly complex geopolitical landscape.”

xAI (Elon Musk’s company) had accepted government requests unconditionally from the start, and was the second company to receive approval for use in classified environments.

Put it all together and a spectrum emerges.

In 2018, “AI should not be used in war” was Silicon Valley conventional wisdom. In 2026, even Anthropic’s position — that it merely asked for two exceptions — is now considered controversial. That’s how fast the Overton window² has shifted.

What the War Revealed About the Supply Chain

The most ironic scene in this entire saga is a separate one.

The day after Trump declared the ban on Anthropic, the US military used Claude in the Iran airstrikes anyway. It was applied to intelligence analysis, target identification, and combat simulation. According to Defense Department officials, replacing Claude would take at least 3 to 6 months. One official compared it to “open-heart surgery.”

This has an important business implication. Once an AI model becomes deeply embedded in a military system, it stops being a simple ‘software license’ and becomes infrastructure — and switching costs become a barrier to change. That’s precisely why the Pentagon was so enraged when Anthropic imposed conditions after the fact. Demanding a change in terms after you’ve already been paid and are deeply embedded in the system is something no client, anywhere, would find acceptable.

At the same time, this is Anthropic’s own paradox. The company built its founding identity on “AI safety,” and its ethical image is a core brand asset. In fact, within a week of this episode, more than a million people a day signed up for Claude, and it topped app store charts in more than 20 countries. It lost a government contract for holding to its principles, but its consumer brand actually got stronger.

Oz’s Lens

Honestly, watching this unfold, I found it hard to simply take one side.

From my own perspective, Anthropic got the business sequencing wrong. Partnering with Palantir in 2024 and embedding itself deep in Defense Department systems, only to publicly demand new conditions in 2026 — both the timing and the method were problematic. Things might have gone differently had it coordinated quietly with the Pentagon. But waging a public campaign through open letters and media interviews, given the nature of the Trump administration, all but guaranteed the predictable backlash it received. That said, since I use Claude myself, maybe I can afford to look at this a little more charitably…

But step back, and the question Anthropic raised still stands. If AI identifies the target, ranks priorities, and supplies the coordinates — and a human only presses the final button — is that really “under human control”? With the kill chain³ compressing from days to seconds, I think it’s a fair question whether a commander can meaningfully review an AI’s recommendation at all.

And this isn’t someone else’s problem. In Korea, the shrinking birthrate keeps eroding military manpower. Deploying AI and robots for border security has already moved past the discussion stage. In this Iran war, watching $1,000 suicide drones disable multi-million-dollar tanks raises fundamental questions about a military structure still centered on self-propelled howitzers and multiple rocket launchers.

Technology has begun putting human lives on the scale. And what tips that scale isn’t AI — in the end, it’s still a human decision.

Closing

Let me sum up the core of this episode in three points.

One, an AI company’s ethical principles turned out to be a peacetime positioning strategy, not a contractual safeguard that holds in wartime. Two, once AI is deeply integrated into a system, it becomes infrastructure — and switching costs carry more negotiating power than principles do. Three, the question of whether it’s fine “as long as AI isn’t the one pulling the trigger” is homework every country will eventually have to answer.

Overseas commentators are already calling this Iran war a kind of expo for global defense AI, and indeed, a striking range of AI technologies — detection, tracking, unmanned systems, reasoning, autonomy — are all being deployed at once. Is a soldier holding a gun somehow fine, simply because a human is holding it? Of course, a beautiful world with no war at all would be the ideal outcome, but that’s too fairy-tale a way of thinking to hold onto realistically — so if we set that aside, I’d just hope that whatever happens ends as fast, and as minimally, as possible.

If you’d like to dig deeper into this topic, I’d recommend the Lawfare legal analysis among the sources below. It gives a meticulous account of just how shaky the legal basis for the supply-chain-risk designation really is.

References & Further Reading

• CNBC, “Pentagon Notifies Anthropic It’s Deemed a Supply-Chain Risk,” 2026.3.5. : The most detailed account of the context surrounding the official supply-chain-risk designation.
• Lawfare, “Pentagon’s Anthropic Designation Won’t Survive First Contact with Legal System,” 2026.3. : A point-by-point legal analysis of the supply-chain-risk designation. Essential reading for understanding the legal dimension of this episode.
• Anthropic official statement, “Where Things Stand with the Department of War,” 2026.3. : Read Anthropic’s official position directly, in its own words.
• OpenAI, “Our Agreement with the Department of War,” 2026.3. : OpenAI’s explanation of the safeguards in its own contract. Worth comparing against Anthropic’s deal.
• CBS News, “Anthropic’s Claude AI Being Used in Iran War,” 2026.3.3. : Reports on how Claude was actually used in the Iran airstrikes.
• Washington Post, “Pentagon Leverages AI in Iran Strikes,” 2026.3.4. : Reports on how Claude was actually used in the Iran airstrikes.
• CNBC, “Google Removes Pledge Not to Use AI for Weapons, Surveillance,” 2025.2.4. : Helpful for understanding the background behind Google’s change to its AI principles. The contrast between the Silicon Valley mood of 2018 and that of 2025 comes through clearly.

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. Maven Smart System: An AI-based decision-support platform Palantir developed for the US Department of Defense. It analyzes vast amounts of data — drone footage, intercepted communications, satellite imagery — in real time to provide intelligence to military commanders. ↩

2. Overton Window: The range of policies or opinions the public considers “acceptable” at a given time. When this window shifts, positions once seen as extreme can become mainstream. ↩

3. Kill Chain: A military term for the full process of finding and identifying a target, then deciding on and executing an attack. As AI becomes involved in this process, cycles that once took days or weeks are shrinking to seconds. ↩