Google, Microsoft, xAI accept CAISI tests

This is the clearest sign yet that frontier-model evaluation is becoming a quasi-gatekeeping layer in the U.S. AI stack, even without a formal licensing regime.

• –CAISI now gets a direct look at unreleased models, including stripped-down versions with safeguards removed, which should surface failure modes earlier.
• –For vendors, the upside is political cover and credibility; the downside is another friction point before launch, especially if testing norms harden into expectations.
• –The move widens the gap between labs willing to subject models to outside scrutiny and competitors that try to ship with minimal oversight.
• –Because the agreement is voluntary, the immediate risk is limited, but the policy direction is clear: more pre-release review, not less.