Anthropic endorsed SB 53 on September 8, 2025, backing Senator Scott Wiener's California bill to regulate the most powerful AI systems built by frontier developers. Rather than prescribing technical safeguards, SB 53 relies on disclosure: large developers must publish safety frameworks, release transparency reports before deploying new models and report critical safety incidents to the state. The bill later became California's Transparency in Frontier AI Act, which took effect on January 1, 2026, and its disclosure model gives developers and enterprise buyers a common yardstick for judging how model providers manage catastrophic risk.

Key Facts#

  • Endorsement: Anthropic announced its support for SB 53 on September 8, 2025, while stating that frontier AI safety is best handled at the federal level.
  • Safety frameworks: Covered developers must publish frameworks describing how they manage, assess and mitigate catastrophic risks, which Anthropic summarized as risks that could foreseeably and materially contribute to a mass casualty incident or substantial monetary damages.
  • Transparency reports: Before deploying powerful new models, developers must publish reports summarizing their catastrophic risk assessments and the steps taken under their frameworks.
  • Incident reporting: Critical safety incidents must be reported to the state within 15 days, and developers must confidentially share summaries of catastrophic-risk assessments for internally deployed models.
  • Accountability: The bill adds whistleblower protections and monetary penalties for developers that break the commitments in their own frameworks.
  • Scope: Coverage depends on training compute, with a threshold of 10^26 FLOPs that Anthropic called "an acceptable starting point." CEO Dario Amodei later wrote that the bill exempts any company with annual gross revenue below $500 million.
  • Status: SB 53 took effect as the Transparency in Frontier AI Act on January 1, 2026, and Anthropic published its compliance framework in December 2025.

What Happened#

Anthropic framed its endorsement as the product of lessons from SB 1047, California's previous attempt at AI regulation, which did not become law. It said Governor Gavin Newsom's Joint California Policy Working Group, made up of academics and industry experts, had recommended a "trust but verify" approach, and that SB 53 put that principle into practice through disclosure requirements instead of the prescriptive technical mandates of the earlier effort.

The company argued that the bill would formalize practices that it and other frontier labs already follow. Anthropic publishes its Responsible Scaling Policy and detailed system cards, and it noted that Google DeepMind, OpenAI and Microsoft have adopted similar frameworks while competing hard. Its central argument was about incentives. Without a law, Anthropic said, labs building ever more powerful models could face pressure to scale back their safety and disclosure programs to keep up with rivals; making disclosure mandatory creates a level playing field. It also welcomed the bill's focus on large companies, with exemptions for startups and smaller firms that are less likely to build the most powerful models.

The endorsement came with a wish list. Anthropic said the compute threshold should be revisited because some powerful models might fall outside it, that developers should disclose more about their tests, evaluations and mitigations, and that regulators should be able to update the rules as the technology advances. It commended Senator Wiener and Governor Newsom and urged California to pass the bill.

Background#

SB 53 shares its core ideas with a framework Anthropic published in July 2025, which proposed that only the largest frontier developers publish secure development frameworks and system cards, with protections for whistleblowers and exemptions for smaller companies. In that proposal, Anthropic floated scoping thresholds such as annual revenue on the order of $100 million or annual R&D or capital spending on the order of $1 billion. Amodei later said Anthropic had proposed an early version of the revenue-based exemption that SB 53 adopted.

State action also reflected a stalled federal debate. The White House's AI Action Plan, released that July, did not set a national transparency standard. Anthropic says it prefers a single federal standard, but it opposed a proposed 10-year moratorium on state AI laws that offered no federal alternative, a provision the Senate removed from the One Big Beautiful Bill by a 99 to 1 vote, according to Amodei. The company's view, stated in the endorsement, was that powerful AI would not wait for consensus in Washington.

Why It Matters for Developers#

Most teams will never be covered by SB 53, which targets the largest frontier developers. The law still changes what they can expect from model providers.

  • Use published frameworks in vendor reviews. Safety frameworks, transparency reports and system cards give procurement and security teams standardized evidence. Keep a registry that links each model version you use to the provider's current framework and system card; the Responsible AI and LLM security guide shows how to build that inventory.
  • Define your own AI incident process. A 15-day reporting clock for frontier developers is a useful benchmark. Decide what counts as an AI incident in your product, who triages it and what telemetry you need, and capture model identifiers and versions with OpenTelemetry as described in the LLM observability guide.
  • Complement vendor tests with your own. Transparency reports describe a provider's catastrophic-risk testing, not how a model behaves in your application. Keep task-specific evaluations, following evaluating AI applications in .NET.
  • Design for a patchwork. Without a federal standard, obligations can differ by state. Centralize logging, redaction and policy checks in IChatClient middleware from Microsoft.Extensions.AI so jurisdiction-specific rules live in one place.

What's Next#

On December 19, 2025, Anthropic published its Frontier Compliance Framework for the new law. It describes how the company assesses and mitigates cyber offense threats, chemical, biological, radiological and nuclear threats, AI sabotage and loss of control, along with its tiered capability evaluations, model weight protection and incident response. The Responsible Scaling Policy remains Anthropic's voluntary safety policy, while the new framework serves as its compliance document for SB 53 and other regulatory requirements.

Anthropic continues to call for a federal transparency law built on public safety frameworks, system cards at deployment, whistleblower protections, flexible standards and coverage limited to the largest developers. Its policy page now says it has also advocated for comparable laws, including New York's RAISE Act and Illinois' SB 315. In 2026 the company went further, arguing in its Advanced AI Framework that transparency alone is no longer sufficient for the most capable models. Whether Congress adopts a national standard, and whether it would preempt laws like SB 53, remains open.

Sources#