Anthropic Warned You About This. Then Sent Six Engineers to Help.

There's a version of the AI safety conversation that sounds principled until you look at who's actually in the room.

According to reporting covered by both TechCrunch and Tom's Hardware, the NSA is reportedly preparing to use Anthropic's Claude — specifically a version called Mythos — for offensive cyber operations. Not defensive. Offensive. As in: cyberattacks. And according to the Tom's Hardware report, roughly half a dozen Anthropic engineers are embedded inside the agency to make that work.

Sit with that for a second.

The Safety Company and the Spy Agency

Anthropic has built its entire public identity around being the careful ones. The adults in the room. The company that thinks about catastrophic risk while the other labs ship fast and apologize later. That positioning has attracted serious investment, serious talent, and serious goodwill from people who find the recklessness of the broader AI industry genuinely alarming.

None of that is fake, exactly. The researchers are real. The concern is probably real. But concern, it turns out, has a carve-out: national security. When the government wants the tool, the tool gets provided — engineers included.

What makes this particularly interesting is the federal angle. TechCrunch noted that the arrangement reportedly exists despite a federal ban on using the AI model maker. The details of that ban aren't fully unpacked in either piece, but the implication is clear enough: there are rules, and then there are exceptions, and the NSA tends to operate in the exception layer.

Ethics as a Feature, Not a Constraint

The uncomfortable truth this surfaces isn't that Anthropic is uniquely compromised. It's that every AI safety framework currently in existence is a set of voluntary commitments made by private companies, enforced by nothing harder than reputation. The moment a sufficiently powerful institution decides the framework is inconvenient, the framework becomes a suggestion.

We've watched this movie with every other technology that mattered. The companies that built the internet's advertising infrastructure didn't set out to build surveillance capitalism. The companies building AI tools for enterprise clients didn't set out to arm intelligence agencies. And yet here we are, with Anthropic engineers reportedly sitting inside the NSA, helping run a model the company publicly positions as safety-first toward ends that are about as far from safety as you can get.

I don't think this makes Anthropic villains. I think it makes them a company that operates in the real world, where government contracts are real money and government relationships are real leverage and saying no to the NSA is a different kind of decision than saying no to a startup that wants to use your API to generate marketing copy.

But it does mean that when Anthropic — or any lab — talks about responsible deployment, you're allowed to ask: responsible to whom, and under what conditions? Because the answer, apparently, is: not under these conditions.

The keys were always going to end up with someone. The only question was who'd be holding them when the story broke.