Anthropic CEO on Open-Weights Models and National Security

Anthropic advocates for specific measures including preventing the sale of powerful chips and chipmaking equipment to China, cracking down on industrial-scale AI model distillation, and implementing mandatory safety testing for all sufficiently capable AI models, both open and closed. These measures aim to address the outlined national security and misuse risks.

Anthropic's CEO explicitly states the company has never advocated for banning open-weights models, viewing them as a public good when devoid of dangerous capabilities. The company believes protectionist bans would not address core national security concerns.

The primary national security concern is that authoritarian governments might develop AI models superior to those in the US, leading to permanent military superiority or deep repression. A secondary concern is the misuse of powerful AI models for cyber or biological attacks, with open-weights models potentially posing a higher risk due to difficulty in monitoring and applying guardrails.

While agreeing that open weights expand access and competition, Anthropic disagrees with assertions that they necessarily ease safeguard development or that broad access always benefits defenders more than attackers. The company worries about potential attacker-defender asymmetry, particularly in biological threat scenarios.