Anthropic Frontier Compliance Framework for California SB 53

Anthropic has released its Frontier Compliance Framework (FCF) to comply with California's Transparency in Frontier AI Act (SB 53), which takes effect on January 1. This framework establishes a public standard for how the company assesses and manages catastrophic risks associated with frontier AI models.

Frontier Compliance Framework (FCF) Overview

The Frontier Compliance Framework (FCF) is the primary document Anthropic uses to meet the legal requirements of SB 53. It details the company's processes for assessing and mitigating specific high-stakes risks, including:

• Catastrophic Threats: The FCF covers cyber offense, chemical, biological, radiological, and nuclear (CBRN) threats, as well as risks related to AI sabotage and loss of control.
• Capability Evaluation: The framework utilizes a tiered system to evaluate model capabilities against these risk categories.
• Operational Security: The FCF outlines the approach to protecting model weights and responding to safety incidents.

Relationship Between FCF and Responsible Scaling Policy (RSP)

While the FCF is the mandatory compliance document for SB 53 and other regulatory requirements, it differs from Anthropic's Responsible Scaling Policy (RSP). The RSP remains a voluntary safety policy that reflects the company's internal best practices and evolves as the AI landscape changes, potentially exceeding current regulatory mandates.

Advocacy for Federal AI Transparency Standards

Anthropic argues that formalizing transparency practices through legislation like SB 53 ensures that safety commitments are not abandoned as competition increases or model capabilities grow. However, the company advocates for a federal standard to ensure national consistency.

Their proposed federal framework includes five core tenets:

1. Public Secure Development Frameworks: Developers must publish how they assess and mitigate serious risks, including CBRN harms and misaligned model autonomy.
2. Deployment System Cards: Public documentation summarizing testing, evaluation procedures, results, and mitigations must be disclosed upon deployment and updated for substantial modifications.
3. Whistleblower Protections: It must be illegal to lie about compliance or punish employees who report violations.
4. Flexible Transparency Standards: Requirements should be lightweight and adaptable to accommodate the evolving nature of AI development.
5. Targeted Application: Regulations should apply only to the largest frontier developers to avoid burdening smaller startups and low-risk models.

Compliance Context

SB 53 establishes the first frontier AI safety and transparency requirements for catastrophic risks in the United States. Anthropic endorsed the law because it balances safety practices, incident reporting, and whistleblower protections with flexibility in implementation and exemptions for smaller companies.