OpenAI vs White House: The New Fault Line in AI Safety Rules

What the White House Wants

Trump’s executive order, signed this week, creates a voluntary framework for evaluating advanced AI models for cybersecurity risks. The lead agency? The National Security Agency.

Voluntary. Military-adjacent. Classified benchmarks. That’s the current White House vision.

What OpenAI Wants Instead

OpenAI’s counter-proposal flips two of those three levers.

First, evaluations should be mandatory — not optional. No lab should get to decide on its own whether its model poses risks. “We don’t think any specific lab should be making that decision unilaterally,” said Chris Lehane, OpenAI’s top policy executive.

Second, the oversight should sit with civilian agencies — specifically CAISI, the Center for AI Standards and Innovation, housed inside NIST at the Commerce Department. OpenAI, Anthropic, and other major labs already have working relationships with CAISI. The NSA has none.

The logic is straightforward: use the infrastructure that already exists, and keep it out of the classified intelligence apparatus.

The Benchmark Problem Nobody’s Talking About

There’s a third wrinkle worth flagging.

The White House order includes a provision to develop AI benchmarking on a classified basis. OpenAI’s Lehane raised an obvious but underappreciated concern: if the capability thresholds that trigger NSA scrutiny are secret, how does any company know when it’s subject to review?

“When do you hit the capability threshold?” Lehane asked. “I think that’s a big part of what the conversation will be.”

It’s a fair question. Regulatory ambiguity at the frontier of AI development isn’t just inconvenient — it’s a genuine governance failure waiting to happen.

Sam Altman Goes to Washington

The timing of OpenAI’s policy paper is not accidental.

CEO Sam Altman arrived in Washington on Wednesday for a round of meetings with White House officials and key lawmakers from both parties. Lehane confirmed Altman would be pushing OpenAI’s preferred framework in those conversations.

The good news for OpenAI: Trump’s order gives agencies 60 days to finalize details. That’s a meaningful window. And congressional allies are already exploring ways to expand CAISI’s authority independently of the executive branch.

OpenAI is playing a long game here — not opposing the White House, but nudging it.

What OpenAI Is *Not* Pushing For

Worth noting: OpenAI isn’t calling for pre-release government approval of new AI models. Lehane explicitly stepped back from that idea, arguing the evaluation framework needs to be built before more aggressive regulations can even be discussed.

“I think you have to do these first steps before you can start to figure out the additional pieces.”

It’s a measured position — and a politically savvy one. Push for mandatory evaluations now. Leave the harder fights for later.

Why This Split Actually Matters

This isn’t just inside-baseball policy drama. The question of who oversees AI model evaluations — and whether that process is voluntary or mandatory, civilian or military, transparent or classified — will shape how every major AI lab operates in the US for years.

If the NSA ends up running a classified, voluntary benchmarking process, the practical result is minimal accountability and maximum opacity. If CAISI gets the mandate and the resources, you get something closer to a functioning, transparent safety infrastructure.

One of those outcomes is significantly better for the broader AI ecosystem. OpenAI knows which one it prefers.

The 60-day clock is ticking. Watch CAISI — it’s quietly becoming the most important acronym in American AI policy.