Microsoft, Google and Elon Musk's xAI have signed a voluntary agreement to share unreleased versions of their frontier large-language models with a federal evaluation programme run out of the National Institute of Standards and Technology, the agency announced on Tuesday. The arrangement, which Tuesday's announcement and a follow-up briefing in Washington on Friday have now fleshed out, gives the new US Centre for AI Standards and Innovation a fixed window of pre-release access in which to test models for cybersecurity, biosecurity and chemical-weapons-related risks.
The Centre, known as CAISI, is the successor to the AI Safety Institute that Joe Biden's administration created at NIST in November 2023. The Trump White House renamed and re-scoped the body in February but largely retained its evaluation methodology, including the practice of red-team probing for capability uplift in narrow biological and chemical-weapons synthesis tasks. The voluntary programme is its first concrete piece of structured industry co-operation.
The agreements are not legally binding and do not give CAISI the power to delay or block a release. The signatories have agreed to allow CAISI a thirty-day window before public release for "frontier" models — defined by training compute above a $10 billion threshold and by capability scoring above a fixed threshold on the Centre's evaluation suite — and a shorter ten-day window for major fine-tunes of existing models.
OpenAI and Anthropic have signed separate, more comprehensive memoranda with CAISI that include longer pre-release access windows and structured access to model weights for evaluation purposes. Meta, which has long argued that its commitment to open weights makes pre-release access redundant, has not signed an agreement of either form. The company released the seventy-billion-parameter Llama 5 family on Wednesday under its existing community licence.
Senator Mark Warner, who chairs the Senate Intelligence Committee, called the agreements "a meaningful first step but obviously not the last word". A bipartisan bill that would give CAISI subpoena authority for frontier-model weights has cleared the committee but is unlikely to reach a floor vote before the August recess, given the crowded national-security legislative calendar.