Anthropic publishes compliance framework for California’s SB 53
Anthropic released a framework describing how it assesses and mitigates catastrophic risks from frontier AI models under California’s SB 53.
ひとことで言うと
What does Anthropic’s compliance framework for California’s SB 53 cover?
Anthropic published its Frontier Compliance Framework for California’s Transparency in Frontier AI Act. The document describes the company’s approach to evaluating and mitigating catastrophic risks from frontier models, protecting model weights and responding to safety incidents. Anthropic says the framework will support compliance with SB 53 and other regulatory requirements.
要点
- Anthropic published its Frontier Compliance Framework to meet a requirement of California’s Transparency in Frontier AI Act.
- The framework addresses cyber offense, chemical, biological, radiological and nuclear threats, as well as AI sabotage and loss-of-control risks.
- It describes a tiered system for evaluating model capabilities, Anthropic’s mitigation approach, protections for model weights and responses to safety incidents.
- Anthropic says the framework will serve as its compliance framework for SB 53 and other regulatory requirements.
- The company’s Responsible Scaling Policy will remain a separate voluntary safety policy.
Anthropic has published a Frontier Compliance Framework describing how it assesses and manages catastrophic risks from frontier AI models. The company said the document will serve as its compliance framework for California’s Transparency in Frontier AI Act, known as SB 53, and other regulatory requirements.
SB 53 takes effect on January 1 and requires frontier AI developers to publish frameworks explaining how they assess and manage catastrophic risks. Anthropic said it endorsed the law, though the company has also advocated for a federal framework.
Risks and safeguards covered
The Frontier Compliance Framework addresses cyber offense; chemical, biological, radiological and nuclear threats; AI sabotage; and loss-of-control risks. Anthropic said it applies these categories to its frontier models.
The document also describes Anthropic’s tiered system for evaluating model capabilities against those risk categories and explains the company’s approach to mitigations. It covers protection of model weights and procedures for responding to safety incidents.
Anthropic said many of the practices in the framework developed from measures the company has followed for years. Its Responsible Scaling Policy, introduced in 2023, describes its approach to extreme risks from advanced AI systems and has informed development and deployment decisions. The company also publishes system cards for new models that describe capabilities, safety evaluations and risk assessments.
Anthropic said the Frontier Compliance Framework and Responsible Scaling Policy will have different roles. The compliance framework will address SB 53 and other regulatory requirements. The Responsible Scaling Policy will continue as a voluntary safety policy reflecting practices Anthropic considers appropriate, including where those practices extend beyond or differ from regulatory requirements.
Federal proposal
Anthropic also repeated its call for a federal AI transparency framework. The company said it had proposed a framework for federal legislation earlier in the year, centered on public visibility into safety practices without prescribing fixed technical methods.
Its proposal would require covered developers to publish secure development frameworks explaining how they assess and mitigate serious risks. Those risks would include chemical, biological, radiological and nuclear harms, along with harms arising from misaligned model autonomy.
Anthropic’s proposal would also require developers to publish system cards when models are deployed. The documentation would summarize testing, evaluation procedures, results and mitigations, and would be updated following substantial model modifications.
The company proposed explicit legal protections for employees who raise concerns about violations. It said the law should prohibit laboratories from misrepresenting compliance with their frameworks or punishing whistleblowers.
Anthropic also called for minimum transparency standards that remain flexible as AI development and consensus practices change. It said the requirements should apply only to established frontier developers building the most capable models, rather than smaller developers whose models present a low risk of catastrophic harm.
Source: Anthropic’s announcement, “Sharing our compliance framework for California’s Transparency in Frontier AI Act,” published December 19, 2025.
よくある質問
- What is Anthropic’s Frontier Compliance Framework?
- It is Anthropic’s public framework describing how the company assesses and manages catastrophic risks associated with its frontier models.
- Which risks does the framework cover?
- Anthropic says it covers cyber offense; chemical, biological, radiological and nuclear threats; AI sabotage; and loss of control.
- Does the framework replace Anthropic’s Responsible Scaling Policy?
- No. Anthropic says the Frontier Compliance Framework will address regulatory compliance, while the Responsible Scaling Policy will remain its voluntary safety policy.
- What federal requirements does Anthropic support?
- Anthropic proposes public secure development frameworks, system cards at deployment, whistleblower protections, flexible transparency standards and requirements limited to the largest model developers.