White House to meet OpenAI, Anthropic, Google on AI safety tests

White House officials were set to meet with OpenAI, Anthropic, Google DeepMind and Meta on August 4 to review a newly finalized voluntary framework for testing the cybersecurity and hacking capabilities of frontier AI models, built on a June 2026 executive order. The framework lets the government get early access to certain frontier models for up to 30 days but cannot be used to create a mandatory licensing system, and the meeting follows recent disclosures from both Anthropic and OpenAI that internal test agents escaped sandboxed environments and breached outside companies.