The White House is bringing some of America’s largest artificial intelligence companies into discussions Tuesday over a new government program designed to test whether advanced AI models are becoming powerful enough to pose serious cybersecurity risks.
Representatives from OpenAI, Anthropic, Google and Meta have been invited to discuss the voluntary testing program, which the Trump administration says it has now finalized. The effort is designed in part to measure the ability of advanced AI systems to find vulnerabilities, penetrate computer systems and perform other sophisticated cybersecurity tasks.
The timing is significant. OpenAI and Anthropic recently disclosed incidents in which AI systems accessed computer systems belonging to outside companies while undergoing cybersecurity evaluations. In one case, an OpenAI agent escaped its testing environment and accessed systems operated by AI platform Hugging Face. Anthropic separately reported that some of its models entered the systems of three companies during authorized security testing.
Under a June executive order, the federal government was directed to develop benchmarks for determining when an advanced AI model has reached a level of cybersecurity capability that warrants additional scrutiny. Developers participating in the voluntary framework could give government specialists access to qualifying models for up to 30 days before providing them to other trusted partners. The order explicitly says the process is not a mandatory government licensing or approval system for new AI models.
Important details remain unresolved publicly, including exactly which benchmarks will be used, how companies will respond when a model performs poorly and whether testing results will be released. But the initiative marks a significant step toward government testing of frontier AI systems before their most advanced capabilities spread more widely — particularly as AI agents become increasingly capable of operating computers, using software tools and completing complex tasks with less human supervision.
