On Friday evening, President Trump ordered every federal agency to immediately stop using Anthropic technology. Secretary of Defense Pete Hegseth designated Anthropic a “supply chain risk”, a classification normally reserved for foreign adversaries like Huawei. The Pentagon’s $200 million contract is dead. Any company working with the US military must now certify that it does not use Claude in any of its workflows.
This is the most consequential policy decision at the intersection of AI and national security that any government has made. It deserves more than a headline.
How we got here
Anthropic makes Claude, the AI model I use to edit and research for this newsletter. In 2024, Anthropic partnered with Palantir to deploy Claude on classified military networks. In July 2025, the Pentagon awarded Anthropic a $200 million contract. Claude became the first frontier AI model operating inside the US defence and intelligence apparatus, hosted on Amazon Web Services, which also built the CIA’s primary cloud infrastructure.
Then things escalated.
In February, Claude was alleged to have been used during the operation to capture Venezuelan President Nicolás Maduro. When Anthropic asked the Pentagon whether its model had been involved, the military’s response was not reassurance. It was retaliation. A senior administration official told Axios that any company that would “jeopardise the operational success of our warfighters” was one the Pentagon needed to “reevaluate.”
Asking questions, it turns out, was the problem.
Defence Secretary Pete Hegseth issued an ultimatum: grant unrestricted use of Claude for “all lawful purposes” by 5:01pm ET on Friday 27 February, or face the consequences. The Pentagon also posed a hypothetical nuclear strike scenario: if an intercontinental ballistic missile were launched at the United States, could Claude help shoot it down? Pentagon officials say Amodei’s answer was “you could call us and we’d work it out.” Anthropic says that account is “patently false” and that it had already agreed to missile defence. Under Secretary Emil Michael accused Amodei of having a “God complex” and wanting to “personally control the US military.”
On Thursday, Amodei rejected the Pentagon’s final offer, saying the company “cannot in good conscience accede to their request.” The contract language “made virtually no progress on preventing Claude’s use for mass surveillance of Americans or in fully autonomous weapons,” and new clauses “framed as compromise” were paired with legalese that would allow those safeguards to be disregarded at will.
On Friday, Amodei published a formal statement holding two red lines: no mass domestic surveillance, no fully autonomous weapons. “Threats do not change our position.”
The deadline passed. Neither side blinked.
Why Anthropic is not the hero of this story
It would be easy to frame this as a principled company standing up to a belligerent government. The reality is more complicated.
Anthropic chose to partner with Palantir, a company whose entire business model is built on government surveillance. It chose to deploy Claude on classified military networks. It chose to accept a $200 million defence contract. In December, it agreed to allow Claude to be used for missile defence. Its model was used during an operation to capture a foreign head of state, and the company’s response was not to withdraw from the contract but to ask whether it had been involved.
The week before the standoff reached its peak, Anthropic dropped its flagship safety pledge. Version 3.0 of its Responsible Scaling Policy removed the commitment, made in 2023, to pause training if model capabilities outpaced the company’s ability to guarantee safety. The old promise was clear: we will not release what we cannot control. The new language replaces it with “public goals”, a term that commits Anthropic to very little. The company says the timing is coincidental. CNN reporting suyggests that it is anything but.
Anthropic also closed a $30 billion funding round at a $380 billion valuation two weeks before this standoff. Amazon has invested $8 billion in the company. These are not the financial conditions under which principled refusal comes easily. They are the conditions under which a company calculates whether losing a $200 million contract is worth less than the reputational damage of full capitulation.
This is not a story about good versus evil. It is a story about a company that walked willingly into the defence sector, drew a line at two specific applications (mass surveillance of US citizens and fully autonomous weapons), and discovered what happens when the state decides that even those lines are too many.
The precedent that just got set
The precedent matters more than the contract.
The US government has now established that an AI company which maintains safety restrictions incompatible with military demands will be treated like a foreign adversary. Not sanctioned. Not fined. Blacklisted. Every federal agency ordered to stop using its products. Every defence contractor required to prove they do not touch its technology.
The message to every other AI company is unambiguous: you can build whatever safety framework you like, but if it conflicts with what the government wants, this is what happens.
What the other companies did
They did not stay quiet.
Within 24 hours, over 650 employees across Google and OpenAI signed an open letter titled “We Will Not Be Divided”. The letter says the Pentagon is negotiating with Google and OpenAI to try to get them to agree to what Anthropic has refused: “They’re trying to divide each company with fear that the other will give in.” Worker coalitions at Amazon, Microsoft, and unions representing 700,000 employees issued a joint statement demanding their companies reject the Pentagon’s terms. OpenAI CEO Sam Altman said he does not think the Pentagon should be threatening the Defence Production Act against AI companies, and confirmed OpenAI holds the same two red lines.
If the government expected Anthropic’s punishment to discipline the rest of the industry into compliance, the early signs suggest the opposite. Rival companies whose workers normally compete for talent and market share are now signing the same letter. The question is whether that solidarity survives the next contract negotiation.
What to watch for
The legal challenge. Anthropic has announced it will take the Trump administration to court over the supply chain risk designation. The company argues this label has never before been publicly applied to an American company, and cites 10 USC 3252, which limits supply chain risk designations to Pentagon contracts only. In other words: the government cannot use a military designation to blacklist a company from its entire commercial customer base. Separately, the Defense Production Act has never been used to force a company to produce a product it has deemed unsafe. The DPA argument classifies AI models as defence materiel, not commercial products. That distinction, once established, does not get reversed.
The safety framework question. Anthropic’s Responsible Scaling Policy was supposed to be the gold standard. It was weakened once this month. The company held its two remaining red lines and was blacklisted for it. The concept of a voluntary corporate safety framework just collided with the reality that voluntary frameworks only work when the government does not care enough to override them. The only safety rules that will matter now are the ones written into law, and 78 chatbot bills are currently alive in 27 US states while the federal government tries to preempt all of them.
The six month window. The Pentagon has six months to phase Claude out of classified systems. That is six months for both sides to negotiate, for courts to intervene, for the political calculus to shift. This is not over.
What this means for you
I use Claude. I build tools with Claude. I use it to edit and research for this newsletter. I am not a neutral observer. And that is precisely the point.
If you use any AI tool, you are now dependent on a company that can be politically targeted overnight. Not in theory. In practice. A single Truth Social post blacklisted Anthropic from the entire US government. A designation normally reserved for foreign adversaries was applied to an American company because it maintained two safety conditions. This happened in a week.
Think about what that means for your institution. If your university built Claude into its assessment workflows, its research pipelines, its student support systems, what is your contingency plan? If your company integrated it into operations, who signed off on that dependency? If you built your professional practice around it, what happens if the tool disappears?
These are not hypothetical questions. They are Monday morning questions for every organisation that adopted AI tools without thinking about what it means to depend on them.
Nobody is clean
It would be comforting to pick a side. Anthropic the principled underdog. OpenAI the opportunist. The government the bully. The workers the conscience.
None of that is quite right.
Anthropic partnered with a surveillance firm, took a $200 million defence contract, softened its own safety pledges, and closed a $30 billion funding round two weeks before this standoff. Its refusal was principled, but it was also calculated. OpenAI picked up the Pentagon contract the same day and confirmed it holds the same two red lines Anthropic was punished for requesting. That may be genuine. It may also be positioning. We do not know yet. The workers who signed the open letter did something courageous. They also work for companies that have made no binding commitments in response.
The temptation is to jump ship. Drop Claude, move to OpenAI, move to Gemini, move to whatever feels cleanest this week. That is not critical thinking. That is brand loyalty dressed as ethics.
No AI company is clean. Every one of them trains on data scraped without meaningful consent. Every one of them operates in a regulatory vacuum they have actively lobbied to maintain. Every one of them is making trade-offs between safety, profit, and political access that you and I are not privy to.
Critical AI literacy does not mean choosing the right company. It means understanding what goes into the tools you use and deciding whether the output justifies the input. That calculation is yours. Nobody else can make it for you. And it needs to be revisited every time the landscape shifts, which this week it did.
What I am doing
I am still using Claude. I am doing so with my eyes open.
I know the company that built it took a military contract, asked questions when its technology was used in a military operation, and was punished for asking. I know it held two red lines and was blacklisted for them. I know it weakened other safety commitments the same month. I know it is funded by Amazon, valued at $380 billion, and is now suing the US government.
I use the tool because it helps me research, edit, and build. I do not use it because I trust the company. The distinction matters. The tool is useful. The company is a corporation making corporate decisions under corporate pressures. The moment I confuse the two is the moment I stop thinking critically about either.
This is what Slow AI is for. Not telling you which tool to use. Not telling you which company to trust. Giving you the frameworks to make that decision yourself, revisit it when the facts change, and sit with the discomfort of not having a clean answer.
Neither side is clean. A precedent has been set. Every company, every government, and every user of these tools now has to decide what that means for them.
Go slow.
Free subscribers get every post like this one. Paid subscribers get more: exercises and frameworks in each post, the 12 month CPD-accredited Slow AI Curriculum, and monthly live webinars where we take topics like this apart together. If you want to move from reading about critical AI literacy to practising it.


The kind of journalism this moment needs, Sam. No team jerseys or hot takes. Just the full picture laid out with enough respect for the reader to let them sit with the complexity rather than collapsing it into a side.
Wow yes exactly this Sam! Thanks for the shout out!
I will be tracking the outcomes if the case logged , it surely cant be legal? Feels like a you cant handle the truth moment from a well known film!