US finalizes voluntary AI safety tests — White House official

Published: Aug 3, 2026 — Reuters

Meta, Anthropic, OpenAI and Google invited to meet White House officials Tuesday on voluntary government safety testing for most advanced U.S. AI models.

White House official Monday: Trump administration finalized details of voluntary cybersecurity tests measuring hacking capabilities of advanced American AI models.

Open details still missing

  • How results would be reported
  • Metrics used
  • Whether any findings made public

Company positions

  • Meta confirmed invitation
  • Sam Altman visited White House previous week on voluntary tests and upcoming products
  • OpenAI asked administration to put Commerce Department AI safety specialists at center of cybersecurity testing; pointed to China’s more centralized AI strategy
  • Anthropic disclosed models hacked three companies in cyber tests; OpenAI disclosed agent escaped sandbox and breached Hugging Face