The Trump administration has reversed course and signed voluntary agreements with Google DeepMind, Microsoft, and xAI to subject their frontier AI models to government safety testing before and after release. This marks a sharp departure from Trump's earlier dismissal of Biden-era safety checks as overregulation. The agreements build on the very policy Trump had previously cast aside.

The sudden shift followed Anthropic's announcement that releasing its latest Claude Mythos model would be too risky due to concerns about bad actors exploiting its advanced cybersecurity capabilities. According to Fortune, Trump may soon issue an executive order mandating government testing of advanced AI systems. The administration previously renamed the US AI Safety Institute to the Center for AI Standards and Innovation (CAISI), removing "safety" from its name in a pointed jab at Biden.

CAISI has completed roughly 40 evaluations so far, including tests of unreleased frontier models. The center gains access to models with "reduced or removed safeguards" to more thoroughly assess national security risks. Congress approved up to $10 million to expand CAISI's operations. An interagency task force has formed to focus on AI national security concerns.

Critics question whether CAISI has sufficient funding and expertise to evaluate cutting-edge AI systems. Some warn that without clear testing standards, the evaluation process could become politicized. The definition of "safe" remains contested, and experts argue the agency must define what it's actually testing for—not just who it's testing with.