Trump administration reverses course on AI safety, signing voluntary testing agreements with Google DeepMind, Microsoft, and xAI after Anthropic withheld Claude Mythos model

The Trump administration signed agreements with Google DeepMind, Microsoft, and xAI to run government safety checks on frontier AI models before and after release. The Center for AI Standards and Innovation (CAISI) has completed about 40 evaluations to date, including those of models not yet released. When testing, CAISI gains access to models with 'reduced or removed safeguards' to evaluate national security-related capabilities and risks.

This reverses Trump's earlier dismissal of Biden-era voluntary safety checks. Trump had rebranded the US AI Safety Institute to CAISI, removing 'safety' from the name. According to White House National Economic Council Director Kevin Hassett, Trump may soon issue an executive order mandating government testing of advanced AI systems prior to release.

Critics warn that CAISI has not specified the testing standards it will use for evaluations, and that the process risks politicization without clear definitions of what 'safe' means. Cornell University experts note that CAISI may lack funding and expertise to evaluate frontier models directly, and that an independent audit system modeled after the IRS might create stronger accountability than voluntary government review.