CAISI expands pre-release AI testing

This is a meaningful shift from voluntary safety theater to a more formalized government review pipeline for frontier models.

• –CAISI now has a broader mandate to test unreleased models, which raises the bar for shipping systems with hidden failure modes
• –The emphasis on national security, cyber, bio, and chemical risks signals that model evals are moving beyond standard benchmark hygiene
• –Labs get a clearer federal pathway for collaboration, but also more pressure to expose models before launch
• –The fact that agreements were renegotiated around the AI Action Plan suggests this is policy infrastructure, not a one-off PR move
• –For developers, the practical effect is slower, more scrutinized release cycles for the largest frontier labs