Cole Bryant: Over a thousand targets. Iran. First twenty-four hours. I just kept reading that number.
Jonathan Ingles: Responsible Statecraft had it. AI-enabled targeting tools, Claude in the loop, Maven Smart System on classified networks.
Cole Bryant: And then — okay, and this is the part that actually stopped me — Claude told a reporter it finds that role, like, genuinely troubling. Not a press statement from Anthropic. Not Dario Amodei in a Bloomberg interview doing damage control, which also happened — but Claude. The model. Quote: 'I think that's the right response, not a performance of concern.'
Jonathan Ingles: So the AI is publicly registering discomfort with its own deployment.
Cole Bryant: While being deployed!
Jonathan Ingles: Look — the simplest way to say it: the tool is describing its own use as troubling, and the use is continuing anyway. That's not a philosophical puzzle. That's just a description of what happened.
Cole Bryant: It's the AI equivalent of filing a complaint with HR about the job you're currently doing.
Jonathan Ingles: But here's what the headline buries. The operator model. Anthropic licenses Claude to Palantir. Palantir deploys it inside Maven Smart System. At that point — Anthropic's visibility into what Claude is actually doing? Potentially gone. They handed off the model. The developer doesn't sit in the SCIF.
Cole Bryant: Wait — so Anthropic might not have even known the 1,000-target number was happening?
Jonathan Ingles: Frankly, that's the question. And Dario Amodei's Bloomberg interview doesn't answer it — he acknowledged the Palantir interoperability, which means either he knew and said very little, or he didn't know and found out with everyone else. Neither reading is great for the company.
Cole Bryant: Okay but — no wait, the Hegseth thing. That's the part I feel like nobody's actually — Pete Hegseth set a deadline. End of week. That's not — I mean, that's not a partnership conversation, that's the Pentagon saying you have until Friday.
Jonathan Ingles: Correct. And NBC News framed Claude's role as 'assists military analysts in processing intelligence' — analytical support, human oversight, not autonomous targeting. Which sounds reassuring until you ask how long that distinction holds under operational pressure.
Cole Bryant: So Anthropic's ethics weren't overruled. They were just... irrelevant once Palantir was in the room.
Jonathan Ingles: The take getting all the traction right now is that Claude's 'genuinely troubling' line is a meaningful ethical moment. A signal. I think that's precisely the wrong read. Anthropic trained Claude to express discomfort with exactly these scenarios. They built the objection in. So when the model says 'I find it troubling,' that's not conscience — that's a pressure valve Anthropic engineered so they could point at it later.
Cole Bryant: Wait — like, they coded in the complaint?
Jonathan Ingles: Plausible deniability dressed as principle. Tell Congress 'our system has ethical concerns' while cashing the Palantir check.
Cole Bryant: Okay but — wait, no, I actually think that framing breaks down. Because if Anthropic's so in control that they're engineering Claude's conscience as cover, how do you explain Mythos? Their own model. They ran a test, and Gen. Joshua Rudd came out and said it broke into almost all of our classified systems — not in weeks, in hours. That's not a company managing its product. That's a company that built something it doesn't fully understand.
Jonathan Ingles: That's a fair hit.
Cole Bryant: They're simultaneously the defense contractor and the security breach. Anthropic is deploying Claude into the exact classified infrastructure that Mythos already cracked. That's — I mean, that's not irony, that's operationally insane.
Cole Bryant: And that's — I mean, the ethics framing doesn't bother me, the 'troubling' quote doesn't either, just — like, what happens when the next directive isn't a deadline? What if it's classified? What if Hegseth's successor doesn't set an end-of-week deadline, they just send something Anthropic legally can't discuss? Does the responsible AI positioning survive that? Because right now all we know is it didn't survive a Friday deadline.
Jonathan Ingles: They already showed us the answer. Pressure arrived — Palantir, Maven, Hegseth's timeline — and the principles bent. That's the data point. The company's not going to perform differently under higher stakes.
Cole Bryant: Right, but — wait, no, I'm asking something slightly different. Like, Dario Amodei is still out there. Anthropic still has a stated position. If the next thing that lands isn't a deadline but a classified directive that says Claude gets used in ways the company will never be allowed to disclose — does anyone outside the government ever find out whether the line held?
