CAISI Widens Pre-Release Model Testing

This is not licensing yet, but it’s the first credible step toward a softer pre-clearance regime for frontier models. Once a small number of labs treat government evaluation as a normal release checkpoint, the line between “technical review” and “approval gate” gets very thin.

• –CAISI says the agreements are voluntary and aimed at information-sharing, not mandatory permission to ship
• –The leverage comes from access: if the government gets unreleased models regularly, it can shape release norms long before Congress writes a formal law
• –The China comparison is fair on process, but the U.S. framing is narrower: national security, cyber risk, and measurement science rather than content control
• –For developers, the biggest risk is scope creep from targeted security testing into broader capability review, especially as models become more agentic and harder to sandbox
• –The fact that CAISI says it has already completed 40+ evaluations suggests this is becoming an operating system for frontier-model oversight, not a one-off experiment