Anthropic publishes framework to comply with SB 53 on AI | Keryc
California is rolling out one of the first laws that require transparency and safety practices for so-called frontier AI. Anthropic has shared its Frontier Compliance Framework (FCF) to explain how it assesses and manages catastrophic risks, and positions it as its guide to comply with SB 53.
What does that mean for you? It promises more visibility into how powerful systems are tested and protected — without turning transparency into mere rhetoric.
What SB 53 requires and why it matters
What do we mean by "frontier AI"? These are the most powerful systems — the ones that could have wide-ranging impacts if they fail or are misused.
California’s law, effective January 1, forces developers of these systems to be more transparent about how they evaluate and mitigate catastrophic risks. It’s an early attempt at setting real expectations for safety.
It’s not just paperwork: SB 53 requires actual safety practices, incident reporting, and whistleblower protections. But it also aims for technical flexibility and exemptions for small companies. The goal? Avoid making transparency a box-checking exercise that only exists on paper.
What Anthropic publishes in its Frontier Compliance Framework (FCF)
In the FCF, Anthropic explains how it evaluates and mitigates risks like cyberattacks, chemical, biological, radiological and nuclear (CBRN) threats, AI-enabled sabotage, and loss-of-control scenarios. It also details:
A tiered system to assess model capabilities against those risk categories.
Mitigation approaches and testing applied at each tier.
Measures to protect model weights and critical intellectual property.
Incident response procedures for security events.
Much of this isn’t new for them: since 2023 Anthropic already had a Responsible Scaling Policy (RSP) and publishes system sheets when it releases new models. The difference now is these practices shift from voluntary to mandatory for developers covered by SB 53.
What Anthropic proposes for a federal standard
Anthropic argues this moment shows the need for a federal framework to align practices nationwide. Its core proposals for a federal law include:
Publishing a secure development framework that explains how severe risks are evaluated and mitigated, including CBRN and autonomy failures.
Publishing system cards at deployment with test results and mitigations.
Protecting whistleblowers and banning labs from lying about compliance.
Keeping transparency standards flexible and lightweight so they can evolve.
Limiting requirements to the largest developers and highest-risk models, so startups aren’t suffocated.
Does that make sense? Yes. If you want public trust, you need real visibility — not just promises.
What this means for industry and for you
For industry: expect more public documentation about tests and mitigations, and stronger incentives to formalize safety practices.
For startups: the law aims not to impose unnecessary burdens, but the debate over where to draw the line will continue. You’ll want to watch how regulators define "high risk."
For society: there’s a better chance to know what precautions are taken in systems that could cause massive harm. Does that mean the risk disappears? No. It means more eyes on processes, and mechanisms to report and fix failures.
Looking ahead
SB 53 is a practical milestone: it pushes good practices out of internal documents and into public view. The logical next step is a federal standard combining transparency, source protection, and technical flexibility.
That would help avoid regulatory gaps between states and set clearer expectations for developers and regulators. The remaining question: can we balance public safety, innovation, and an environment where startups can grow? That’s the discussion ahead, and we need technical, regulatory, and public voices at the table.
