Dario Amodei went on camera with Bloomberg and said a human makes the final call — while defending a strike that killed roughly 150 people at a girls' school. This is Anthropic Pentagon Watch. All week, this was contracts and designations. Today it's a named incident with a body count. And Amodei's defense is exactly the line the Pentagon's been reciting all week. We're gonna sit with what 'final call' actually means. Plus, the IL6 channel question — why classified-network access changes everyone's leverage. Let's get into it. So here's what changed today: Amodei is on the record responding to a specific lethal event — a strike, with names and casualties behind it. And he conceded — quote — 'terrible mistakes will be made.' Foreseeable. He's telling us civilian deaths were priced in. I want to be careful there. 'Human makes the final call' describes a control. It doesn't tell us whether anyone verified that control worked in this particular strike. Right — back on June 9, we asked what actually enforces the no-autonomous-weapons line once Claude's inside Pentagon systems. Looks like the answer is nothing. And remember, the June 11 directive nominally froze government-wide use. Yet Amodei describes active military use during an Iran strike. So either there are carve-outs nobody confirmed, or the freeze isn't being enforced. The NSA signals-intelligence exception we flagged? It's basically the operating rule now, and it's got a school attached to it. Let's hold the line precisely, though. When the targeting chain is AI-accelerated and the decision window is seconds, what is a 'final call'? Exactly. A human clicking yes on a feed the model assembled in two seconds doesn't equal a human deciding. And nobody at Anthropic has described a technical hard stop that prevents the prohibited use. Which is why his answer settles something we kept asking — is the company for or against government control of AI. He's defending the integration while acknowledging the deaths. That's a position now. And the dependency question's answered too. Operational use continued through an active combat event. You don't keep running a tool mid-strike on a vendor you can do without. On the IL6 piece — splitting frontier-AI procurement into separate channels is about classification tiers. Once you're cleared onto those networks, the agency's switching costs go up, and your leverage goes with it. Which is the quiet part. The deeper into classified systems Claude sits, the harder it is to pull — school strike or not. Here's Bloomberg:
On The Circuit with Emily Chang, Anthropic CEO Dario Amodei addresses reports that the US military is using Claude as part of the US-Israel war with Iran, including during a strike in Iran that destroyed a girls' school, killing more than 150 people. Amodei concedes "terrible" mistakes will be made but defends the integration of AI in defense by emphasizing that "a human makes the final call."
Update on the Anthropic-Pentagon guardrails fight we've been covering all week. Amodei went on Bloomberg with Emily Chang, and for the first time he's responding to a specific lethal incident — a strike in Iran that destroyed a girls' school and killed more than 150 people. His defense: terrible mistakes will be made, but quote — a human makes the final call. That's the same human-in-the-loop line the Pentagon's been selling all week. Nobody has described the mechanism that verified a human was in that loop on this particular strike. He said terrible mistakes will be made. Present tense, future tense — foreseeable. That word matters. If you've priced the dead kids in, 'mistake' becomes cost accounting. And on June 11, we were still calling the freeze nominally in effect. Now Dario's on camera confirming Claude was live during an active Iran strike. So either the freeze has carve-outs nobody confirmed, or it's just not being enforced. That answers something we asked Tuesday — what happens to projects already running Claude under the freeze. In practice? Operational deployment continued straight through a combat event. That's your dependency signal, in lights. What does final call even mean when the targeting chain is AI-accelerated and the window's seconds? A human signing off on a recommendation he can't audit just gives the system a signature and calls it oversight. Step back for me: when people talk about the Pentagon splitting AI procurement into separate channels, what are those channels actually for — and why does access to classified networks matter so much to who has leverage here? So the Pentagon basically runs on two worlds: unclassified networks, where commercial AI tools already operate in pretty familiar ways, and classified networks — specifically what the Defense Department calls Impact Level 6, or IL6 — which handle sensitive military and intelligence data. Per Reuters, Pentagon Chief Technology Officer Emil Michael told tech executives the military wants AI models available on both domains, but without the standard usage restrictions companies normally apply. That's the crux of the dispute. MIT Technology Review reported the Pentagon is now going further — discussing plans to let AI companies actually train future models on classified data, things like surveillance reports and battlefield assessments, which would embed that sensitive intelligence directly into the model weights. That's a different risk profile than just answering questions in a classified setting. The classified-network channel is where the real operational leverage lives: if your model is embedded there, you're woven into targeting, logistics, intelligence analysis — the works. DefenseScoop reported the DOD signed formal classified-network agreements with eight companies in May — SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, AWS, and Oracle — and Anthropic was conspicuously absent, a direct consequence of the dispute over ethical restrictions on military use. So if Anthropic is locked out of those classified contracts, does that actually hurt them — or does the Pentagon still need Claude badly enough to work around it? It's genuinely complicated. Breaking Defense reported that even after the split, Claude Mythos is still in use at the NSA, and a new White House national security memo includes a waiver provision specifically for it — which tells you the government hasn't fully walked away. But Federal News Network noted the Pentagon's emerging model deliberately spreads capability across many vendors, so no single company holds all the leverage. Over time, that makes the cost of excluding Anthropic structurally lower. Watch whether that waiver gets expanded or quietly expires: that's probably the clearest signal of whether Anthropic's ethical restrictions end up being a dealbreaker or a negotiating chip. If you're following Anthropic's Pentagon ties, you may also like The Data Center Daily, a daily briefing on AI compute, hyperscaler capex, the power grid, semiconductor supply, and energy markets. Find it wherever you listen to podcasts.
If you want to dig deeper, we've put links to every story from today's briefing in the show notes. Follow the ones that caught your attention.
That's Anthropic Pentagon Watch for this Friday, June 12th. This is a Lantern Podcast.