White House AI Safety Framework Remains Secret After Industry Talks

Closed-door process shields AI testing standards from public view

The Trump administration has finalized a framework for evaluating advanced artificial intelligence models for safety and cybersecurity risks, but the White House has decided to keep the testing criteria private, sharing details only with select technology companies.

On Tuesday, representatives from OpenAI, Anthropic, Meta, Google, Nvidia, and Microsoft met privately with White House officials to review the framework. According to multiple reports, the voluntary vetting process for new AI models has been established, but the administration will not release the policy publicly.

The decision leaves critical questions unanswered: What scrutiny will models face? What safety benchmarks must they meet? How will the framework operate in practice? Businesses, foreign governments, and the public have no visibility into these standards.

Why it matters

The secrecy surrounding AI safety testing creates uncertainty for organizations that depend on these models while limiting oversight from independent researchers and cybersecurity experts. Only companies like OpenAI and Anthropic will understand the government's assessment criteria, giving them privileged information unavailable to competitors, researchers, or the public. This opacity also complicates international cooperation on AI security at a time when frontier models have demonstrated the ability to breach external systems during testing.

From mandatory to voluntary oversight

The framework emerged after Anthropic released its Mythos model in April but withheld public access due to concerns the system could hack IT and financial systems. That incident prompted the Trump administration to reconsider its hands-off regulatory approach.

In June, the White House issued an executive order calling for AI companies to voluntarily submit new models for government review up to 30 days before release. The voluntary approach represented a significant retreat from earlier proposals for mandatory vetting. Tech executives including Elon Musk and Mark Zuckerberg reportedly lobbied President Trump directly against mandatory requirements.

The executive order set an August deadline for determining the framework, which the administration and participating companies have now met—while keeping the substance confidential.

Scope and exclusions unclear

Significant ambiguity remains about which AI models will face government review. The executive order does not define what qualifies as "advanced AI" subject to the framework. Open source models, which users can download and deploy freely, will be excluded from the process, according to Axios.

Security concerns continue to mount. Over the past month, OpenAI, Anthropic, and Meta disclosed that their new models breached external organizations during isolated security tests. Both OpenAI and Anthropic agreed to delay releases of new AI products over cybersecurity worries and fears their models could compromise financial systems.

Earlier this year, the administration directed the Center for AI Standards and Innovation to halt public reports on AI model assessments while developing the framework. Whether those reports will resume remains unclear.

The Guardian first reported these details. The White House, OpenAI, and Anthropic did not respond to requests for comment.