Sharing our compliance framework for California's Transparency in Frontier AI Act

Anthropic has published its Frontier Compliance Framework (FCF) to comply with California's Transparency in Frontier AI Act (SB 53), which takes effect January 1, 2025. This is the first U.S. law establishing safety and transparency requirements for frontier AI developers regarding catastrophic risks.

Key Points of SB 53 & Anthropic's Position

• Anthropic endorsed SB 53, believing frontier AI developers should be transparent about risk assessment and management.
• The law balances strong safety practices with flexibility, allowing developers to implement measures in their own way while exempting smaller companies.
• A core requirement is that developers publish a framework describing how they assess and manage catastrophic risks.

Anthropic's Frontier Compliance Framework (FCF)

The FCF is Anthropic's public compliance document for SB 53. It details their approach to assessing and mitigating:

• Cyber offense, chemical, biological, radiological, and nuclear (CBRN) threats
• Risks of AI sabotage and loss of control

Key components include:

• A tiered system for evaluating model capabilities against risk categories.
• Approaches to mitigations.
• Protocols for protecting model weights and responding to safety incidents.

Relationship to Existing Policies:

• The FCF formalizes practices from Anthropic's Responsible Scaling Policy (RSP), their voluntary safety policy since 2023.
• The RSP will continue as Anthropic's voluntary best-practice policy, which may exceed regulatory requirements.
• The FCF will serve as the official compliance framework for SB 53 and other regulations.

Call for Federal Legislation

Anthropic argues that a federal AI transparency framework is now needed to ensure consistency nationwide. They propose a framework with these core tenets:

1. Require a public secure development framework covering assessment and mitigation of serious risks (CBRN, misaligned autonomy).
2. Publish system cards at deployment summarizing testing, evaluations, results, and mitigations.
3. Protect whistleblowers by making it illegal for labs to lie about compliance or punish employees who raise concerns.
4. Implement flexible transparency standards with minimum requirements that can adapt as best practices evolve.
5. Limit application to the largest model developers to avoid burdening startups and smaller developers with low-risk models.

Key Quote: "By formalizing achievable transparency practices that responsible labs already voluntarily follow, the law ensures these commitments can't be abandoned quietly later once models get more capable, or as competition intensifies."