Pre-Deployment Information Sharing: A Zoning Taxonomy for Precursory Capabilities

Early warning systems can help catch dangerous AI capabilities before they become critical threats. The paper proposes breaking down high-impact capabilities into smaller "precursory capabilities" that appear earlier in development, allowing researchers and regulators to detect progress toward dangerous abilities long before they reach critical "red lines."

Information should be shared proactively during development, not just after deployment. The authors recommend that AI companies share information about precursory capabilities as soon as they're discovered through internal testing, rather than waiting until a system is ready to be released to the public—enabling earlier intervention opportunities.

Specialized AI Safety Institutes should serve as trusted intermediaries for sensitive capability information. Rather than sharing dangerous capability details broadly, the paper recommends that appointed AI Safety Institutes receive and coordinate this information, acting as a central point for tracking progress and enabling coordinated responses.

Information security must increase as capabilities approach dangerous thresholds. The framework proposes that as precursory capabilities move closer to potentially harmful endpoints, the information about them should receive stronger protections—potentially including classification or restricted access—to prevent misuse while still enabling legitimate safety oversight.

A structured "zoning" approach can clarify when and how to share information about AI risks. By creating a taxonomy that maps capability development stages, the framework provides concrete guidance for when different actors should be informed and what level of detail is appropriate at each stage.