OpenAI's head of hardware resigned on Saturday over the company's decision to give the Pentagon access to its AI models — without first defining the guardrails around surveillance and autonomous weapons. It's the most visible internal dissent OpenAI has faced in months, and it lands in the middle of a messy industry-wide battle over who controls AI in classified environments.
What Happened
Caitlin Kalinowski, who joined OpenAI in November 2024 from Meta — where she led the team building the Orion AR glasses — announced her resignation on LinkedIn on March 7. Her reason: OpenAI rushed into a deal with the Department of Defense without locking down the rules first.
"AI has an important role in national security," she wrote. "But surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got."
In a follow-up post on X, she clarified the target: "My issue is that the announcement was rushed without the guardrails defined. It's a governance concern first and foremost. These are too important for deals or announcements to be rushed."
She was explicit that this wasn't a personal fallout — she has "deep respect" for Sam Altman and the team. This was about principle.
The Pentagon Deal Nobody Agreed On
The OpenAI-Pentagon agreement was announced around March 1, and it allows OpenAI's models to run on classified cloud networks. On paper, OpenAI says there are two hard limits: no domestic surveillance, no fully autonomous weapons. But Kalinowski's point is that those limits were declared publicly without internal governance frameworks to enforce them — a press release, not a policy.
The backstory makes this more interesting. The Pentagon first approached Anthropic. Anthropic tried to negotiate real safeguards into the agreement — preventing use in mass surveillance and autonomous lethal systems. The Pentagon refused, then designated Anthropic a supply-chain risk for pushing back. Anthropic is fighting the designation in court.
Within days, OpenAI moved in to fill the gap.
Whether that makes OpenAI pragmatic or reckless depends on who you ask. Kalinowski, from inside the building, picked a clear side.
Why This Matters Beyond OpenAI
Kalinowski's resignation is a signal, not just a personnel move. A few things worth watching:
The Anthropic contrast is stark. Anthropic walked away from a deal that didn't meet its safety bar. OpenAI took the deal. The AI safety community will read this as OpenAI crossing a line Anthropic refused to cross — and will debate whether Anthropic's stance cost them or defined them. Microsoft, Google, and Amazon have quietly confirmed they'll continue making Claude available to non-defense customers.
Internal dissent at OpenAI is rare and notable. The company has survived a boardroom coup, mass departures of safety researchers, and the pivot from nonprofit to for-profit with relatively contained internal pushback. A named, principled resignation from a hardware exec — someone without an obvious ax to grind about AI safety politics — is different. Kalinowski wasn't a policy researcher. She built things. Her quitting sends a different kind of signal.
The governance gap is real. OpenAI's statement says the deal "makes clear our red lines." But red lines without enforcement mechanisms are just public commitments. If the Pentagon asks OpenAI to do something that edges toward those lines, who decides? What's the process? Kalinowski's core complaint — governance, not people — is the exact question OpenAI hasn't answered publicly.
This is what AI commercialization in defense actually looks like. Not a clean handshake deal with agreed-upon rules. A rushed announcement, internal protest, a competitor pushed out for asking the hard questions first. The market pressure to be the AI partner of record for the US military is enormous — OpenAI's valuation depends partly on the assumption that it will win large government contracts. That pressure doesn't slow down for policy frameworks.
What's Next
OpenAI's Pentagon deal is done. Kalinowski is out. The real question now is whether other employees raise similar concerns internally — or quietly leave — and whether Anthropic's court challenge to the supply-chain designation succeeds.
If Anthropic wins, it reframes the dynamic: companies that hold the line on safety terms can't be punished for it. If they lose, it's a green light for the Pentagon to pressure every frontier AI lab into compliance by threatening procurement blacklists.
Either way, the AI-military relationship just got a lot harder to ignore. And the internal cost at OpenAI just became visible.
FOLLOW NeuralWire on X for daily AI signal — what matters, why it matters, what to do about it. →