In the span of five days, an AI company refused the Pentagon's demands, got blacklisted by the President, watched its competitor swoop in with its own deal, dropped its own flagship safety pledge, got community-noted on X, became the subject of a 300-person employee revolt across rival companies, hit #1 on the App Store, and then had its technology used in combat anyway.
Is your head-spinning yet? Cause ours is!
Anyway, this is the story of how we got here.
Part 1: How Claude Ended Up on the Battlefield
To understand what happened last week, you need to understand what happened last year.
In November 2024, Anthropic partnered with Palantir and Amazon Web Services to deploy Claude on classified military and intelligence networks. It was a big deal. Palantir's platform is the backbone of US intelligence operations, and getting a frontier AI model cleared for classified work at DISA Impact Level 6 (the second-highest security tier) was something no other AI company had accomplished.
By June 2025, Anthropic had launched Claude Gov, a version tailored for government and national security workflows. A month later, in July 2025, the Department of Defense awarded Anthropic a contract worth up to $200 million to "prototype frontier AI capabilities that advance US national security." OpenAI, Google, and xAI received similar contracts for unclassified work; Claude was the only model on classified networks.
By late 2025, Claude was embedded deeply across US military operations. Intelligence assessments. Target identification. Combat simulations. Operational planning. According to multiple reports, Claude had become effectively indispensable to the Pentagon's AI infrastructure.
Then came the Caracas raid.
On January 3, 2026, the US military launched an operation to capture Venezuelan President Nicolás Maduro. The Wall Street Journal reported that Claude was used during the operation through Palantir's classified platform. Between 83 and 100 people were killed during the raid.
According to Axios, a senior Pentagon official claimed an Anthropic executive reached out to a Palantir executive to ask whether Claude had been used, "in such a way to imply that they might disapprove of their software being used, because obviously there was kinetic fire during that raid, people were shot." Anthropic flatly denied this, saying it had not "discussed the use of Claude for specific operations with the Department of War."
Regardless of who's telling the truth, the Caracas raid was the catalyst. It forced a question that had been simmering since the partnership began: what happens when a company that prohibits its technology from being used for autonomous weapons discovers that technology is being used in lethal operations?
Part 2: The Ultimatum
The Pentagon had been pushing all four AI labs (Anthropic, OpenAI, Google, xAI) to agree to an "all lawful purposes" standard for military use. OpenAI, Google, and xAI had agreed to lift their consumer-facing guardrails for Pentagon work. Anthropic had not.
Anthropic's position was narrow but firm: Claude could be used for any lawful military purpose except two things. No mass domestic surveillance of Americans. No fully autonomous weapons without human oversight.
That's it. Not a blanket refusal to work with the military. Not a pacifist stance. Just two red lines.
According to Axios's reporting, the Pentagon viewed this as unworkable. A senior administration official argued there was "considerable gray area" around what would and wouldn't fall into those categories, and that "it's unworkable for the Pentagon" to operate under restrictions defined by a private company rather than the law.
The Pentagon's frustration was also personal. Dario Amodei had spent months taking increasingly aggressive public stances against the administration's AI policies, criticizing Trump's attempt to ban state-level AI regulation and donating $20 million to a group supporting congressional candidates who pledged to push for AI safety laws.
On Tuesday, February 25, Defense Secretary Pete Hegseth met with Amodei at the Pentagon. Six senior defense officials attended, including the department's top lawyer. The ultimatum: grant the military unrestricted access to Claude by Friday, or face consequences. Those consequences included being declared a "supply-chain risk" under federal procurement law, or compulsion under the Defense Production Act, a Cold War-era law that gives the president authority to force private companies to prioritize national security needs.
On the same day, Elon Musk's xAI signed a deal with the Pentagon to deploy Grok on classified networks, agreeing without restriction to the "all lawful purposes" standard. Claude was no longer alone behind the classified wall.
Part 3: The Safety Pledge That Disappeared
Here's where the timeline gets uncomfortable.
On the same Tuesday that Hegseth was issuing his ultimatum, Anthropic published a blog post announcing it was overhauling its Responsible Scaling Policy. The RSP was Anthropic's signature safety framework, first introduced in 2023, built around one central promise: Anthropic would never train or deploy an AI system unless it could guarantee that its safety measures were adequate.
That promise was now gone.
The new policy replaced the hard trigger with something softer. Anthropic would only "delay" development if it considered itself to be leading the AI race and believed the risks of catastrophe to be significant. Instead of categorical commitments, the company would publish "Frontier Safety Roadmaps" and "Risk Reports" at regular intervals. Instead of hard lines, public goals.
Anthropic's chief science officer Jared Kaplan told TIME: "We didn't really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments... if competitors are blazing ahead."
Anthropic said the change was unrelated to the Pentagon dispute. The timing, let's just say, raised eyebrows.
As MIT physicist Max Tegmark told TechCrunch:
"All four of these companies have now broken their own promises. First we had Google; this big slogan, 'Don't be evil.' Then they dropped that. Then they dropped another longer commitment that basically said they promised not to do harm with AI. They dropped that so they could sell AI for surveillance and weapons. OpenAI just dropped the word safety from their mission statement. xAI shut down their whole safety team. And now Anthropic, earlier in the week, dropped their most important safety commitment."
The broader point Tegmark made is worth sitting with: the AI industry spent years lobbying against binding regulation, promising to govern itself. Now, in Tegmark's words, "we have less regulation on AI systems in America than on sandwiches." And the companies that promised self-governance are, one by one, loosening the self-imposed rules they pointed to as evidence that regulation wasn't needed.
Part 4: The Line
On Thursday, February 27, Dario Amodei published Anthropic's formal response.
"We cannot in good conscience accede to their request," he wrote. He acknowledged the Pentagon's authority to make military decisions, but maintained that two specific applications crossed ethical boundaries that Anthropic would not help enable.
According to Axios, the contract language the Pentagon sent overnight included loopholes that would have allowed the mass surveillance and autonomous weapons safeguards to be overridden at the government's discretion. Anthropic viewed this as a "compromise" that wasn't really a compromise at all.
Within hours, the hammer came down.
President Trump posted on Truth Social: "THE UNITED STATES OF AMERICA WILL NEVER ALLOW A RADICAL LEFT, WOKE COMPANY TO DICTATE HOW OUR GREAT MILITARY FIGHTS AND WINS WARS!" He directed every federal agency to "IMMEDIATELY CEASE all use of Anthropic's technology," with a six-month phase-out period.
Defense Secretary Hegseth followed up by designating Anthropic a "supply-chain risk to national security" under 10 USC 3252, a statute normally reserved for foreign adversaries like Huawei and ZTE. It was the first time the designation had ever been applied to an American company.
Anthropic vowed to challenge the designation in court, calling it "retaliatory and unprecedented." CEO Dario Amodei told CBS News: "Disagreeing with the government is the most American thing in the world."
Part 5: OpenAI Steps In
Hours after Anthropic was blacklisted, Sam Altman posted on X that OpenAI had reached an agreement with the Department of War to deploy its models on classified networks.
"AI safety and wide distribution of benefits are the core of our mission," Altman wrote. He claimed the deal included the same two red lines as Anthropic's (no mass surveillance, no autonomous weapons) plus a third: no AI-powered social credit systems. He also said OpenAI would deploy "FDEs" (field deployment engineers) and use cloud networks only, with "technical safeguards to ensure our models behave as they should."
The post got community-noted almost immediately. Government officials contradicted Altman's framing, saying OpenAI had agreed to allow its models to be used for "all lawful purposes," the exact standard Anthropic had refused. Community note readers added that this "could allow for uses Anthropic refused to engage in, namely mass surveillance tools and weapons systems with no human oversight."
The optics were brutal. Sam Altman had spent the previous 24 hours saying he shared Anthropic's red lines. He told CNBC that morning it was important for companies to work with the military "as long as it is going to comply with legal protections" and "the few red lines" that "we share with Anthropic." An internal note to staff on Thursday evening said OpenAI was seeking to negotiate a deal that excluded surveillance.
Then, hours later, the deal was signed.
On X, critics were unsparing. "You backstabbed Anthropic," one prominent account wrote. "Actions speak louder than words." Others pointed out that the contract language appeared to contain nothing preventing constrained mass surveillance, surveillance of non-US citizens, or autonomous weapons not already covered by existing law.
To his credit, Altman held an AMA on X the next day. He said OpenAI had told the DoW, both before and after signing, that Anthropic should not be designated a supply-chain risk. He said part of the reason OpenAI acted quickly was "in the hopes of de-escalation." He also raised a genuine philosophical question: "There is more open debate than I thought there would be, at least in this part of Twitter, about whether we should prefer a democratically elected government or unelected private companies to have more power."
That is, in fairness, the right question.
But the timing of the answer, signing a deal within hours of a competitor being blacklisted for holding the line that OpenAI publicly claimed to share, left a lot of people unconvinced. As TechCrunch's interview with Max Tegmark noted: "Last night, Sam Altman came out and said he stands with Anthropic and has the same red lines. I admire him for the courage of saying that." The editor's note added: "Hours after the interview, OpenAI announced its own deal with the Pentagon."
Part 6: The $110B Elephant
This all happened against the backdrop of OpenAI's other major announcement that same week: a $110 billion funding round at a $730 billion pre-money valuation.
The round included $50 billion from Amazon ($15B deployed immediately, $35B conditional on milestones including a potential IPO), $30 billion from Nvidia (largely in the form of compute capacity commitments on next-generation Vera Rubin chips), and $30 billion from SoftBank (described by some observers as a letter of intent rather than committed capital).
The Pentagon deal and the funding round are not unrelated. OpenAI is preparing for an IPO, likely later this year. It needs to demonstrate that it can work with governments, win enterprise customers, and scale infrastructure. A $200M classified contract isn't transformative for a company valued at $840B, but the signal it sends to investors, regulators, and sovereign wealth funds absolutely is.
Anthropic, for its part, had raised $30 billion at a $380 billion valuation just weeks earlier. Anthropic also has its own IPO timeline, and its business model (direct-to-enterprise rather than consumer-first) could cost it customers who do business with the government, which to be fair, is A LOT of customers.
Part 7: The Industry Reacts
Since all this started, over 300 Google employees and 60+ OpenAI staffers signed an open letter published at notdivided.org urging their own companies to hold the same ethical line Anthropic had drawn. The letter called on all AI companies to refuse unrestricted military access.
This matters because it came from inside the house. Google and OpenAI employees were publicly pushing back against their own employers' willingness to comply with the Pentagon's terms, just as those employers were positioning themselves to take over Anthropic's classified contracts.
Meanwhile, the public reacted with a surge of support for Claude. Claude hit #1 on the US App Store, overtaking ChatGPT. Katy Perry posted a screenshot of her Claude Pro subscription with a heart emoji. A "Cancel ChatGPT" movement gained more traction on Reddit and X. Anthropic's daily signups tripled since November; paid subscribers doubled in 2026. Anthropic even launched a new feature to let you import your memories from ChatGPT into Claude; nice way to take advantage of the free press!
Getting blacklisted by the President of the United States turned out to be pretty good marketing! Thanks, Chief Storyteller Pete Hegseth!
Part 8: The Kicker
And then came the part that made the whole week feel like satire.
On Saturday, March 1, the Wall Street Journal reported that the US military used Claude during air strikes on Iran, just hours after Trump had ordered all federal agencies to stop using the technology.
Apparently, CENTCOM had Claude running intelligence assessments, identifying targets, and simulating combat scenarios. The same AI model the President had publicly banned was being used in active military operations the same day.
The Pentagon's explanation was practical: Claude is so deeply embedded in classified systems that removing it immediately would disrupt ongoing operations. The six-month phase-out exists for exactly this reason. The alternative AI models (OpenAI, Google, xAI) aren't yet integrated into classified networks at the same level.
But the optics are devastating. The government banned a company for refusing to allow unrestricted military use of its AI; then used that AI in a military operation before the ink on the ban was dry.
What This All Means
There are several ways to read this week.
The optimistic read: A company drew an ethical line, and the market rewarded it. Claude's App Store surge, the employee open letter, and the public backlash against OpenAI suggest that consumers and workers value principled stances, even when they come at a cost. Anthropic bet that principles could be a business model, and so far, that bet is paying off.
The cynical read: This was a vendor dispute that got political. The Pentagon wanted unrestricted access. Anthropic wanted to keep its terms-of-service restrictions. The government escalated because Amodei had been publicly antagonizing the administration for months. OpenAI stepped in because it needed the government relationship for its IPO narrative. Everyone is acting in their own interest and dressing it up in principles.
The structural read: This was inevitable the moment AI companies started lobbying against regulation while simultaneously promising to govern themselves. As Tegmark put it, if the companies had pushed for binding AI safety laws when they had the chance, they wouldn't need to rely on their own terms of service to draw lines. The government would already be drawing them. Instead, the industry has corporate amnesty, no regulatory framework, and a set of increasingly powerful tools that the government can demand access to with no legal guardrails beyond existing law.
The truth is probably all three.
What's undeniable is that the question at the center of all this isn't going away: when AI becomes critical military infrastructure (and it already has), who decides the rules? The companies that build it? The government that uses it? The laws that govern both?
Right now, there's a vacuum where those rules should be. Anthropic tried to fill it with its own policies. The government said that's not your job. And the rest of the industry quietly signed whatever the government asked for, while publicly claiming to share Anthropic's values.
The week ended with Claude being used in a war zone, banned and essential at the same time. If there's a better metaphor for where AI policy stands in 2026, we haven't found it.
