This is the kind of story that sounds abstract until you realize it’s really about who gets to set the rules for powerful tech: the people building it, or the government buying it.
Anthropic is in a fight with the Pentagon, and it just took a bad turn for the company. Based on public reporting, a federal appeals court upheld the Pentagon’s decision to blacklist Anthropic from military contracts. The reason isn’t that Anthropic is accused of being tied to a foreign government. It’s that Anthropic wouldn’t remove its own restrictions on how its AI can be used. The court ruling matters beyond one company because it appears to broaden how much power the executive branch has to label a U.S. company a national security risk—based on product safety policies, not just foreign affiliations.
If you’re thinking, “Wait, a safety policy can be treated like a security threat?” Yeah. That’s the tension.
On one level, I get the Pentagon’s frustration. If the military wants to use a tool, it doesn’t want a vendor telling it “no” or “only in these situations.” Defense agencies don’t like being boxed in, and they especially don’t like it when the box is built by a private company that answers to its own board, its own lawyers, and its own reputation. From their view, restrictions could mean delays, limits, and uncertainty in situations where they believe speed and flexibility save lives.
But here’s what bothers me: treating a company’s refusal to loosen guardrails as grounds for blacklisting flips the incentive structure in a really ugly way. If the lesson to every AI company is “you’ll lose government business if you keep firm limits,” then the market will start rewarding the most permissive, least restrictive products. And with AI, “permissive” doesn’t just mean more features. It can mean easier misuse.
Imagine you run an AI company and you’re trying to decide whether to block certain uses—say, anything that helps target people, manipulate public opinion, or build weapons. You know you’ll take heat either way. If you set limits, some customers complain. If you remove limits, you may end up enabling harm and wearing the blame later. Now add a third pressure: the government might punish you for having those limits at all. A lot of leaders will make the pragmatic choice and quietly loosen the rules.
That’s not “national security.” That’s national security theater mixed with procurement leverage.
The most important detail here is the precedent: executive authority to classify U.S. companies as national security risks based on product safety policies. That’s a huge hammer to introduce into the relationship between government and domestic tech. It’s one thing to say “we don’t like your pricing” or “your product doesn’t meet our needs.” It’s another to say “your safety stance makes you a threat.”
Because once you accept that logic, it won’t stop at one AI company. Today it’s “remove usage restrictions.” Tomorrow it’s “build in access for us,” or “don’t log that,” or “don’t publish that safety research,” or “we need this capability and we need it quietly.” You don’t have to be paranoid to see how that goes.
And look, I’m not pretending Anthropic is some flawless hero here. Companies love to talk about safety, but they also love growth, and they also respond to pressure. Sometimes “restrictions” are real, sometimes they’re fuzzy policy language that can change when money talks. The public usually can’t see the private negotiations, the carve-outs, the exceptions. So I’m not handing out moral gold stars.
Still, I’d rather have a world where an AI company can say, “We’re not comfortable enabling this,” without being treated like an enemy.
There’s also a bigger, uncomfortable question underneath: who should decide what’s “acceptable” AI use in war and intelligence? If you’re the kind of person who thinks the military should have every tool available because adversaries won’t hold back, then you’ll see Anthropic’s restrictions as naive at best, dangerous at worst. If you think the danger of AI misuse is exactly why we need hard lines, you’ll see the Pentagon’s stance as coercion dressed up as security.
Both sides can tell a story where they’re protecting people.
But I keep coming back to consequences. If government pressure becomes the standard way to break safety limits, the long-run outcome isn’t just one company losing contracts. It’s a race to build systems that are easier to bend, easier to aim, and harder to question. It also pushes safety debates out of public view and into closed rooms where the public can’t weigh trade-offs.
At the same time, if AI companies get total control to refuse whole categories of use—including defense—then we’re effectively outsourcing national defense choices to private policy teams. That’s not obviously good either. It can turn into a world where a handful of executives decide what a country is allowed to do, based on their own risk tolerance and brand concerns.
So yes, I’m worried about the court upholding a blacklist tied to safety policies. I think it rewards the wrong behavior and normalizes a dangerous kind of pressure.
But I’m also not totally sure what a healthy alternative looks like when the stakes are real and other countries won’t play by the same rules.
If the government can punish companies for keeping AI guardrails, and companies can unilaterally deny tools the government wants, who should have the final say on where the line is?