Safety plan
Published frontier safety framework. Write down how you assess and mitigate catastrophic risk, publish it, and follow it.
What it does
The lab publishes a document that says which dangerous abilities it tests its models for, what it does when a model crosses a line, and how it guards its models. Then it has to follow what it wrote.
Applies to: Large frontier developers (revenue above $500M).
- Publish a frontier AI framework covering capability thresholds, mitigations, pre-deployment review, third-party assessment, weight security and incident response (SB 53 § 22757.12).
- Review it at least yearly; publish material changes within 30 days with a justification.
- Do not make materially false statements about compliance with it.
- EU signatories must adopt an equivalent Safety and Security Framework.
Where things stand
Anthropic's Responsible Scaling Policy is on version 3.4. OpenAI's Preparedness Framework is on version 2, with a companion Frontier Governance Framework written to map onto SB 53 and the EU Code. Google DeepMind and xAI publish frontier safety frameworks. The requirement codifies what every covered lab already publishes.
Why it costs almost no lead
Every major lab already publishes one. Legal review and upkeep displace almost no research: under a day.
Biggest unknown: Whether being legally bound to your own framework makes a lab write a weaker one or hold a launch to comply with it.
Why it lowers p(doom) by ~0.05%
Makes each lab commit in public to what it will do if a model turns out to be dangerous.
Labs already publish these. Making them binding helps a little by making it harder to quietly drop commitments under pressure.
The strongest case that it costs more
Anthropic's RSP v3.0 explicitly moved away from hard pre-deployment gates toward public goals it grades itself against. A law that makes a framework binding removes that flexibility. If a lab's own threshold is crossed at an awkward moment, a binding framework could force a delay the voluntary version would not. The cost is real but it is the cost of a threshold, and it is priced under the evaluation and safety-case items.
The debate
For
- California SB 53 requires large developers to publish a frontier AI framework, 2025.
- Seoul Frontier AI Safety Commitments 20 companies including Anthropic, Google, Microsoft, OpenAI and Meta pledged to publish safety frameworks, 2024.
- Anthropic said SB 53 would formalize practices it already follows, 2025.
Against
- Chamber of Progress said SB 53 imposes no meaningful safety duty and is a compliance minefield, 2025.
- OpenAI's Chris Lehane, CTA and a16z lobbied against SB 53 or raised objections, as reported by TechCrunch, 2025.
Sources
- California Bus. & Prof. Code § 22757.12 (SB 53): frontier AI framework
- Anthropic Responsible Scaling Policy, version history: v3.4 effective July 8, 2026
- OpenAI Preparedness Framework v2: April 15, 2025
- OpenAI Frontier Governance Framework: maps Preparedness practices onto SB 53 and the EU Code
- GovAI on Anthropic's RSP v3.0
Rough starting points, not precise forecasts. Lead costs assume China doesn't depend on U.S. models, the case least favorable to safety laws, and count 3 years. On the menu you can change every assumption and put in your own numbers. Last priced 2026-09-26.