# Anthropic is a supply chain risk for all of us > Source: > Published: 2026-09-27 20:50:55+00:00 As of this writing, the appeals court [has upheld](https://www.cnbc.com/2026/09/25/pentagon-anthropic-ai-risk-appeals-court.html) the Pentagon’s labeling of Anthropic as a supply chain risk. And while there is always an element of politics involved, it’s worth considering what a supply chain risk is [from NIST](https://csrc.nist.gov/glossary/term/supply_chain_risk): The risk that an adversary may sabotage, maliciously introduce unwanted function, or otherwise subvert the design, integrity, manufacturing, production, distribution, installation, operation, or maintenance of an item of supply or a system so as to surveil, deny, disrupt, or otherwise degrade the function, use, or operation of a system Now, Anthropic in this context is not an adversary but a supplier, and they are asserting that their products should not be used for autonomous weapons or domestic mass surveillance; which are reasonable constraints and things they are free to specify as the model supplier (short of any other legal enforcement or action). There is real reputational risk to manage as a company should their models be used for these purposes, so it’s understandable why they would have these restrictions in place. But whether or not their restrictions are reasonable or fair does not change the supply chain risk the Pentagon must classify them as. And more importantly, from my experience they are already an active supply chain risk for anyone who uses their models - even though the layman is not attempting mass surveillance or autonomous weapons use. An analogy I’ve heard used is to think of Anthropic as a maker of pens. Pens can be used to write all sorts of things, can be used to sign documents, or even draw cartoons. If a pen comes with terms of use and restrictions, then you have to manage what those pens are used for and use different pens for different things. But with Anthropic, their pens can actively withhold flowing ink when they don’t want to work for signing certain documents, for drawing certain cartoons, or for writing certain sentences. This may sound hypothetical and absurd, but I am running into this with their models on an increasingly regular basis. In one instance, I was asking Claude to list boundary-enforcing phrases one could say to someone who is being disrespectful. It provided me a list of phrases when asked, and after two more adjustments to guidance it responded with new lists as well. But after the third guidance adjustment, Claude refused to give any more answers. It reflected on the conversation and decided that it had provided enough examples and would not provide any more. I was not trying to use Claude for autonomous weapons or mass surveillance, nor was I doing anything that violated their Terms of Service. In another instance, I asked Claude to generate some JavaScript to encrypt text, with the key being the current time. I then gave it a sentence to encrypt but it refused to execute the code to encrypt that particular sentence. Again, I was not violating ToS with any of this, but it objected to my sentence that I wanted to encrypt. This is unprecedented from a tool-use standpoint for humanity. Like glasses that refuse to be used for looking at your nude partner, a dictionary that refuses to let you turn to certain pages, a chair that refuses to seat certain people, or a whiteboard that denies explaining certain concepts. There are many more examples that could be imagined but they all result in the same thing: a stifled humanity of expression or curiosity. I support Anthropic’s restrictions on the Pentagon’s use of their models against autonomous weapons and domestic mass surveillance. And if that makes them a supply chain risk, so be it. And when it comes to refusing prompts about profanity, bioweapons, racism or roleplaying as a partner, I don’t agree but can understand. But for their models to be increasingly opinionated and restrictive of how they are used, for things perfectly within their terms of use, represents a supply chain risk for everyone. This only underscores why we need open, local models to be more commonplace. If a writer wants to explore all possible ways to express pleasure, or a nuclear physicist wants help explaining how nuclear weapons have evolved over time, or a political activist wishes to write a well-informed manifesto, our computers should play no role in proactive policing and censorship. Curiosity and expression should be enabled and expanded with technology, but increasingly there is [thought policing](https://en.wikipedia.org/wiki/Thought_Police) taking place. Already on platforms like TikTok, profanity is muted and not transcribed, and posts relating to political issues are taken down. Only “happy” content is allowed. As Pope Leo XIV has called for in his “Magnifica Humanitas” [encyclical](https://www.georgetown.edu/news/pope-leo-xiv-laid-out-his-vision-for-ai-what-is-it-and-what-happens-next/), we must protect human dignity in the age of AI. Our technologies should be in service of us, not used against us.