Frontier Model Security

As the capabilities of frontier AI models rapidly improve, securing systems is emerging as a core challenge. Anthropic emphasizes that preventing model theft or misuse requires security far exceeding the standards of existing commercial technology.

In particular, advanced AI models, model weights, and the research data underpinning them must be protected. To this end, the AI field should be treated as 'critical infrastructure' to strengthen public-private cooperation, and future compliance enforcement through government procurement or regulation should also be considered.

As a core principle for security, 'two-party control' is proposed. This is designed so that no single individual can have persistent access to the operational environment, meaning a 'multi-party authorization' approach in which, when access is needed for work, a colleague's approval is required and access is granted only for a limited time.

Additionally, secure software development practices compliant with standards such as NIST's SSDF (Secure Software Development Framework) and SLSA (Supply Chain Levels for Software Artifacts) should be spread across the frontier AI environment as a whole.