In this position statement, Anthropic CEO Dario Amodei addresses the controversy surrounding open-weights models and proposes a strategic framework for managing AI-related national security and safety risks. Amodei explicitly rejects the notion that Anthropic supports a blanket ban on open-weights models, arguing that they provide significant public utility, provided they do not possess dangerous capabilities.

Core Concerns

Amodei identifies two primary "nightmare scenarios" that inform his policy recommendations:

1. Geopolitical Superiority: The risk that authoritarian regimes (specifically the CCP) develop AI models superior to those of the US to achieve military dominance or enhance domestic repression.
2. Safety and Misuse: The risk that highly capable models could be weaponized for biological or cyberattacks, compounded by potential alignment failures. He notes that open-weights models are harder to constrain once released, as they lack the centralized guardrails and monitoring capabilities inherent in closed-source, API-driven systems.

Proposed Policy Framework

Amodei argues that protectionist bans on open-weights models are ineffective, proposing instead a three-pronged regulatory approach:

• Supply Chain Restrictions: Preventing the export of high-end compute hardware and chipmaking equipment to China. This strategy relies on the principle that scaling laws dictate performance; by restricting the available compute $C$, the effective capability $E$ of a model is inherently limited by the relationship $E \approx f(C)$, where $f$ represents the scaling function. Restricting access to high-performance chips acts as a direct bottleneck on training state-of-the-art models.
• Mitigation of Industrial Distillation: Amodei identifies "distillation"—the process of training smaller, efficient models on the outputs of larger, frontier-level models—as a significant evasion tactic. Because distillation is compute-efficient, it allows actors to leverage existing state-sponsored frontier models to develop powerful local iterations, effectively narrowing the gap between foreign and domestic capabilities. He calls for targeted legal frameworks to disrupt these operations.
• Mandatory Pre-release Safety Testing: Amodei advocates for a global standard of rigorous, mandatory testing for all "sufficiently capable" models, regardless of whether they are open or closed-source. This testing should specifically evaluate risks related to biological synthesis, cyber offensive capabilities, and alignment. He acknowledges the "attacker-defender asymmetry," noting that the development of biological agents via AI could proceed faster than the development of medical countermeasures, necessitating empirical verification of safety before model release.

Technical and Philosophical Stance

Amodei expresses skepticism toward the argument that open weights inherently favor the "defender" in security contexts. He posits that in fields like biotechnology, the barrier for an attacker to synthesize harmful agents is lower than the operational capacity for a state to defend against them, suggesting that wide access to capabilities may increase net global risk. Consequently, he calls for future research into modular training strategies and robust safety testing protocols as the only viable methods to reconcile the benefits of open-source AI with the requirements of national security.