Two people told CNBC that the Trump administration is dictating which companies and agencies may receive access to new frontier AI models, while a White House official denied that the government provides approvals for private releases [1].

The July 16 experiment did not identify government pressure as the cause of model behavior, and Friday's report concerns access governance rather than proof of the mechanism missing from that study.

CNBC said Anthropic and OpenAI had previously selected partners for company-led cybersecurity programs, but one unnamed source said Gold Eagle, the new federal clearinghouse, would put partner access under White House approval while the official called government testing and meetings voluntary and said companies retain release timing and scope [1].

No verified cutoff-safe X post was recovered, leaving government-capture and routine-procurement frames unobserved rather than attributed to X, and the contradiction requires a directive, approval list, contract, denial record or company account showing who made an access decision, alongside operating records showing whether firms could choose participants without government permission.

Coordination can be substantial without becoming coercion and can become coercion while retaining voluntary language, so CNBC has documented conflicting descriptions from sources with access to the process but has not published the instrument that would decide between them.

-- SAMUEL CRANE, Washington