The companies building the models get to see the security playbook, but the people those models will affect don't.
The Summary
- The White House shared its AI model evaluation framework with OpenAI, Anthropic, Microsoft, and other major labs on Tuesday, but won't release it publicly
- The timing is strange: OpenAI and Anthropic both recently suffered high-profile security breaches that rattled public confidence in AI security
- The administration hasn't explained why a cybersecurity framework meant to protect the public is being kept from the public
The Signal
The Trump administration convened the major AI labs Tuesday for a closed-door review of how they should evaluate their models for security risks. The framework is meant to establish baseline standards for AI cybersecurity. The companies building the most powerful AI systems in the world now know what those standards are. Everyone else is guessing.
Fortune called the secrecy "baffling," and they're right. This isn't classified defense research. It's a framework for evaluating commercial AI products that millions of people already use. The stated goal is to make AI systems safer. But you can't verify safety standards you can't see.
"The companies building the models get the playbook. The researchers, civil society groups, and users who could actually stress-test those standards are locked out."
The timing makes this worse. Both OpenAI and Anthropic recently got hacked. Details are scarce, but the breaches were serious enough to spook the public. Now the administration rolls out a cybersecurity framework in response, shows it only to the companies that got breached, and tells everyone else to trust the process.
This isn't how security frameworks normally work. NIST's Cybersecurity Framework is public. So is their AI Risk Management Framework. When the government sets standards for critical infrastructure or emerging technology, transparency is the norm. It has to be. Public review catches gaps. Independent researchers find edge cases. Sunlight is the actual disinfectant.
Here's what we don't know:
- What evaluation criteria are in the framework
- Whether it's mandatory or voluntary
- How enforcement would work
- Who wrote it and what input they took
The administration hasn't said why the framework is secret, which means we're left to guess. Maybe they think public disclosure would help bad actors. Maybe they're worried about revealing gaps in current AI security. Maybe they just don't want the scrutiny. None of those reasons hold water. Security researchers can't help close holes they can't see. The public can't pressure labs to meet standards they don't know exist.
The Implication
If you're building AI systems or working in AI safety, you're operating in the dark. You don't know if your security measures meet the government's bar because the government won't tell you what the bar is. If you're a researcher trying to audit AI models for security flaws, you can't align your work with official standards because those standards are classified.
Watch for leaks. Someone in that room Tuesday will eventually talk. When they do, we'll learn whether this framework is substantive or theater. In the meantime, assume AI labs are writing their own report cards, and the White House is fine with that.