Executive Summary: The Blind Spot in AI Governance
In this doctrinal paper, Irma Arguello argues that the current global focus on AI safety is dangerously incomplete. While safety efforts aim to prevent unintentional system failures, they leave a massive flank exposed: intentional security threats.
As AI models reach the "open-weight" frontier, they effectively become sovereign tools with no "kill switches," capable of being weaponized by malicious actors. This analysis introduces the Dual-Risk Framework, a strategic approach that draws lessons from nuclear non-proliferation to address both the internal reliability of the system and the external dangers of human misuse.
This research is a cornerstone for policymakers and strategists looking to move beyond technical safety into the realm of global security and strategic stability.
- The limitations of the Safety paradigm:Global AI governance currently focuses on safety—ensuring systems don't cause unintentional harm. However, this is insufficient because it ignores security risks: the deliberate misuse or weaponization of AI by malicious actors.
- The Danger of Open-Weight Diffusion: Releasing model weights is a direct transfer of operational power. Once public, these systems become sovereign and lack "kill switches," enabling actors to replicate or weaponize them beyond oversight.
- Strategic Lessons from the Nuclear Domain:Just as fissile material is the core danger in nuclear weapons, model weights are the "fissile core" of AI. We must protect these core sources of danger rather than offering them to potentially harmful actors at no cost.