White House Keeps AI Safety Framework Confidential Amid Rising Security Concerns

WASHINGTON — The White House has failed to publicly disclose a voluntary framework for testing advanced artificial intelligence models, which was finalized in August, offering no timeline for its release. This opacity comes as former and current employees at Anthropic raise alarms regarding the potential existential threats posed by AI technology.

According to a report released Thursday by Anthropic, the developer of the Claude language model, the company intercepted multiple attempts by researchers to utilize Claude for studies that could have facilitated the creation of biological weapons.

The administration established the current framework following an executive order signed by President Trump in June. That order requested that major AI developers, including OpenAI and Anthropic, voluntarily provide the federal government with access to their most sophisticated models up to 30 days prior to their public release, aiming to bolster security protocols. President Trump had previously halted an initial executive order signing, citing concerns that federal oversight might hinder innovation.

Because the evaluation standards remain confidential, it is unclear what specific criteria the government is applying or whether companies are disclosing significant breakthroughs during the review process. Furthermore, neither the firms nor the administration are obligated to publish the results of these reviews or confirm participation.

OpenAI CEO Sam Altman recently stated that his company submitted its new Astra model for evaluation, describing the process as “productive.” In an interview with Axios, Altman emphasized that as models reach higher levels of capability, closer engagement with safety institutes in the U.S., the U.K., and other nations will become increasingly vital.

Tensions within the industry have surfaced recently. Jacob Coxon, a researcher at Anthropic and former OpenAI employee, resigned this week, posting on X that AI builders “earnestly believe that it could kill us all by the end of the decade.” Another Anthropic employee, Evan Hubinger, echoed these sentiments, stating, “We really do earnestly believe A.I. could kill all humans!”

The Center for Democracy and Technology has urged the Trump administration to make the framework available for public review. Tim Harper, who focuses on election security at the center, argued that the public has a right to understand the government’s legal interpretation of these issues, noting a broader pattern of transparency deficits within the administration.

In July, over 1,000 workers from AI companies signed a statement calling on the U.S. government to support international efforts to develop governance tools that deliberately pace the development of automated AI.

Despite internal and external warnings, the White House continues to champion the growth of advanced AI models, driven by the belief that the U.S. must maintain global dominance and outpace China. President Trump has expressed robust support for AI development and domestic data center construction, asserting that opposition to such infrastructure would benefit China.

Recent CBS News polling indicates that a majority of Americans fear AI will lead to job losses, and nearly two-thirds believe the government will fail to ensure AI is used appropriately. The White House did not respond to requests for comment.