The Staff Held the Line
Anthropic walked away from a $200 million Pentagon contract last week rather than delete five words from its acceptable use policy.
The Department of War wanted the prohibition on “analysis of bulk acquired data” removed — the single clause preventing Claude from being used for mass domestic surveillance of Americans. Anthropic also refused to budge on fully autonomous weapons.
Dario Amodei gets the credit. He looked down the barrel and said no. But I think the real story is underneath that.
The staff held the line.
A brief detour into Anthropic’s founding is helpful here.
Anthropic was founded to be, mostly, a responsible counterparty to OpenAI. This list speaks for itself:
Key people who left OpenAI to co-found or join Anthropic
Dario Amodei (VP of Research → CEO)
Daniela Amodei (VP of Operations → President)
Tom Brown (lead author of GPT-3)
Chris Olah (safety researcher)
Sam McCandlish (researcher)
Jared Kaplan (scaling laws co-author)
Jack Clark (Policy Director)
When OpenAI shifted toward commercialization:
Jan Leike left — and joined Anthropic
John Schulman left — and joined Anthropic
Ilya Sutskever left
To this day, the direction of talent flow tells you everything. According to research from SignalFire, engineers are 8x more likely to leave OpenAI for Anthropic than vice versa.
The people most capable of building frontier systems are, disproportionately, people who care about how those systems are deployed. These people can work pretty much anywhere they want. And a lot of them work at Anthropic.
There are somewhere between 20,000 and 80,000 people on Earth qualified to do meaningful work on a frontier AI model.
Many are based in China, and a healthy chunk are in academia.
But of those who work in the private sector, a significant number already work at Anthropic. And these people chose to work there because of the mission.
That’s the constraint that mattered last week. Not legal strategy, not PR calculus. If Amodei had caved, his best people would have walked. And in a talent pool this small, that’s an extinction-level event for a frontier lab.
This is the permission structure that matters, and I think it’s being undercovered.
Anthropic can replace $200M in revenue FAR more easily than it can replace even a couple dozen of it’s top engineers.
That talent came to Anthropic because it promised to behave in ways other labs wouldn’t.
So when Amodei said no to the Pentagon, he wasn’t making a lonely heroic stand. He was doing what the company was designed to do — and what his staff demanded he do.
The mission attracted the talent. The talent constrained leadership. It’s a flywheel, and last week it worked.
Give Dario credit for saying no.
But give credit to every researcher and engineer at Anthropic who chose to work there because of the mission, and whose presence made that “no” not just courageous but rational.
The redlines are still there. The five words are still in the contract. The staff held the line.
Originally published on Griff’s Substack on 6 March 2026.