The Daiki SB-53 Recipe

A Practical California SB-53 Frontier AI Regulation Compliance Roadmap

The global landscape of artificial intelligence governance, long a patchwork of national strategies, voluntary commitments, and ethical guidelines, is undergoing a structural shift. We are moving from the era of principles to the era of enforcement. Following the comprehensive, risk-based taxonomy of the EU AI Act and the process-oriented framework of ISO/IEC 42001, a new and potent regulatory instrument has emerged from the epicenter of AI development: California’s Senate Bill 53 (SB-53).

SB-53 – the Transparency in Frontier Artificial Intelligence Act (TFAIA) – is the first U.S. law that targets frontier models explicitly, with a focus on catastrophic risk, incident reporting, and standards-based transparency. It represents a paradigm shift in U.S. technology regulation. As the Carnegie Endowment analysis puts it, SB-53 offers “a blueprint for evidence-generating transparency measures” that may shape frontier AI governance for years to come.

For enterprises and developers, this law introduces a non-negotiable regulatory friction that demands a new model of governance—one that is continuous, provable, and deeply integrated into the AI lifecycle. As we have argued in previous Daiki work on the EU AI Act, ISO/IEC 42001, and responsible quantum governance, this path must be built not on reactive checklists but on a proactive, systemic, governance-model-based (GMB) approach.

For business leaders, this shift means AI can no longer be treated as a technical add-on delegated solely to engineering teams: governance around models, data, and downstream use now determines regulatory exposure, resilience, and licence to operate. Robust AI governance is increasingly a board-level question about strategy, accountability, and long-term enterprise value—not merely a compliance checkbox.

Below is a Daiki-style deep dive into:

• what SB-53 does (and does not do),
• how it came about,
• when and to whom it applies, and
• how companies and institutions can build business value through an efficient self-regulatory AI compliance program using Daiki SB-53 Recipe.

1. How we got here: from SB-1047 to SB-53

SB-53 did not emerge in a vacuum. It is the product of a multi-year California debate on how far a state can (and should) go in regulating high-end AI development.

In 2024, California first attempted a frontier-AI safety law with SB-1047, a bill that would have imposed pre-training safety requirements, third-party audits, and strong shutdown obligations on large models. Governor Newsom vetoed SB-1047 as overly burdensome and called for further work, warning that the proposal risked stifling innovation in the state’s core industry.(

Following the veto, he convened the Joint California Policy Working Group on AI Frontier Models, which in 2024 produced the California Report on Frontier AI Policy: a detailed set of recommendations prioritising evidence-generating transparency, incident reporting, and a state-backed compute cluster over direct prohibitions.

SB-53, authored by Senator Scott Wiener, largely implements those recommendations in a narrower, more “procedural” way. Instead of ex ante licensing or kill switches, it focuses on:

• post-deployment transparency,
• catastrophic-risk management, and
• whistleblower protection,

with standards and federal frameworks as reference points rather than hard-coded technical rules.

Governor Newsom signed SB-53 on 29 September 2025, making California the first U.S. jurisdiction with a statute specifically focused on frontier AI models. Most requirements take effect on 1 January 2026, giving frontier developers a short runway to design and publish their frameworks, transparency reports, and incident-reporting channels.

For Daiki clients, the key point is simple: SB-53 is now a live legal obligation for a small but highly visible set of AI developers – and a template others will be asked to emulate.

2. Scope: who is covered, and for which models?

2.1 Frontier models and developers

SB-53 is intentionally narrow. It does not regulate all AI. It targets a specific layer: frontier foundation models, like the ones created by Google, OpenAI, Microsoft, Meta and Anthropic – trained with very large amounts of compute.

The law defines, among others:

• “Foundation model” – an artificial intelligence model trained on a broad dataset, designed for generality of output, and adaptable to a wide range of tasks.
• “Frontier model” – a foundation model trained using more than 10²⁶ integer or floating-point operations (FLOPs), including compute used for substantial fine-tuning, reinforcement learning, or other material modifications.
• “Frontier developer” – any person or entity that has trained or initiated the training of a frontier model meeting that compute threshold.
• “Large frontier developer” – a frontier developer whose annual gross revenue (including affiliates) exceeded USD 500 million in the preceding calendar year.

In practice, external analyses suggest only a small number of models have publicly disclosed training runs above the 10²⁶ FLOP threshold so far, but more are expected to cross it as scaling continues.

SB-53 is not limited to California-incorporated developers. Like privacy and consumer-protection law, it is expected to apply to companies with sufficient nexus to California markets (for example, making frontier models available to users in the state).(JD Supra)

2.2 Catastrophic risk and critical safety incidents

SB-53 is framed around “catastrophic risk” and “critical safety incidents”.

• Catastrophic risk is defined as a foreseeable and material risk that development, storage, use, or deployment of a frontier model will materially contribute to:
  • death or serious injury of more than 50 people, or
  • more than USD 1 billion in damage or loss of property,

from a single incident, through scenarios such as AI-enabled CBRN weapons, autonomous cybercrime and physical violence without meaningful human oversight, or loss of control over the model.

• A critical safety incident includes, among others:
  • unauthorized access to or exfiltration of model weights that results in death or bodily injury;
  • harm from the materialisation of a catastrophic risk;
  • loss of control of a frontier model causing death or serious injury;
  • deceptive model behaviour that subverts developer controls and materially increases catastrophic risk.

The thresholds are deliberately high. As analysis by IAPP notes, SB-53 targets low-probability, high-impact harms (the catastrophic risk that would be classified as ‘unacceptable risk’ in the EU AI Act’s pyramid of criticality) and uses those cases as a driver for evidence-generating transparency, rather than general AI liability or broad licensing.

Here’s a self-contained section you can drop into the blog, plus a clear suggestion on placement.

2.3 What does SB-53 mean for Daiki customers?

SB-53 is narrow by design, but its practical reach is broader than the small set of companies that obviously qualify today. Even if you are not yet a “frontier developer” or “large frontier developer” in the strict statutory sense, SB-53 will influence what your investors, customers, and partners expect from your AI governance. It is therefore useful to distinguish several archetypes among Daiki’s clients.

Startups and scaleups building on frontier models

Most early-stage and growth-stage companies will not cross the 10²⁶ FLOPs threshold themselves. They build on top of APIs and platforms provided by the large labs. Strictly speaking, SB-53 obligations attach first to those upstream frontier developers.

But there are three reasons why startups and scaleups should still care:

1. Contractual back-pressure. As upstream providers implement frontier AI frameworks, incident-reporting channels, and whistleblower protections, they will tend to push analogous obligations down the chain through terms of service, DPAs, and enterprise contracts. Startups that already have a structured AI management system (ISO/IEC 42001-style) will be far better positioned to sign those contracts without disruption.
2. Investor and customer diligence. Investors and enterprise customers will increasingly treat “SB-53-ready” governance as a hygiene factor. Young companies that can show a Daiki-style registry of models, risk assessments, and incident workflows will look more credible than those with ad hoc governance.
3. Forward compatibility with future thresholds. SB-53 allows definitions and thresholds to be updated over time. A scaleup that is “just below” the frontier line today may cross it faster than anticipated. Treating the Daiki SB-53 Recipe as a forward-compatible governance scaffold avoids a disruptive scramble later.

In practice, Daiki will help startups and scaleups:

• maintain a single AI registry that already tracks compute, sector, and geography;
• reuse EU AI Act and ISO/IEC 42001 artefacts as the backbone of their SB-53 story;
• plug in only those SB-53 elements that matter now (e.g., basic incident logging, whistleblower policy) while keeping room to grow into full frontier-developer status.

Large multinationals and hyperscale builders

At the other end of the spectrum are Daiki clients that:

• train large foundation models themselves, and
• have global revenues well above the USD 500 million “large frontier developer” mark.

These entities are the primary target group of SB-53. For them, the law is not a theoretical future constraint; it is a concrete obligation that:

• turns their internal safety framework into a legally relevant document;
• makes transparency reports and incident reporting mandatory for frontier models;
• exposes them to civil penalties if they misstate catastrophic risk or fail to implement the processes they advertise.

For these clients, Daiki’s SB-53 Recipe is intended to be deployed group-wide, not only inside California entities. It provides:

• a frontier model register and compute-threshold tracking across legal entities;
• a structured frontier AI framework that can be published once and then tailored by region;
• linked workflows for transparency reports, OES incident notifications, and whistleblower protection;
• mappings into EU AI Act high-risk and GPAI obligations, ISO/IEC 42001 controls, and NIST AI RMF.

The operational question is no longer “Does SB-53 apply?” but “How do we avoid creating five parallel governance stacks?” Daiki’s answer is a single, standards-centric AI management system with SB-53 as one regulatory overlay.

EU companies and institutions doing business in California

EU-based companies and institutions – including universities, research consortia, and public bodies – face a dual reality:

• On one side, they must comply with the EU AI Act and sectoral EU rules.
• On the other, if they train frontier models and make them available to users in California, they are likely to fall within the scope of SB-53 just as U.S. developers do.

For these organisations, SB-53 does three things:

1. Adds a jurisdictional overlay on top of their existing EU AI Act implementation.
2. Raises the bar on evidence-generating transparency – system cards, incident logs, and whistleblower channels that can be inspected by California authorities.
3. Increases the value of a harmonised AI management system that can serve both Brussels and Sacramento from one governance backbone.

Daiki’s platform already treats the EU AI Act and ISO/IEC 42001 as first-class reference points. The SB-53 Recipe simply activates an additional layer: frontier-specific templates for risk, transparency, incidents, and whistleblowers, without duplicating artefacts.

EU clients that do not train frontier models but consume frontier APIs will still feel SB-53 indirectly, via:

• contractual clauses from U.S. providers;
• due diligence questions from Californian customers;
• expectations for incident reporting and safe-use policies when their applications are deployed in the state.

Public institutions and research labs

Finally, SB-53 is highly relevant to public research labs and universities, both in California and abroad, that:

• participate in state-backed initiatives such as CalCompute; or
• run large-scale open science frontier models.

These institutions may lack the commercial compliance infrastructure of big tech, but are increasingly expected to meet similar safety and transparency standards when their models cross frontier thresholds. Daiki’s SB-53 Recipe gives them:

• a lightweight SB-53 governance skeleton they can adapt to academic and public-law settings;
• a way to demonstrate to funders and partners that their safety work is systematic, not ad hoc;
• a bridge between open science values and the risk-based obligations in SB-53 and the EU AI Act.

3. Core obligations under SB-53

SB-53’s obligations fall into five main buckets: frontier AI frameworks, transparency reports, incident reporting, whistleblower protection, and enforcement.

3.1 Frontier AI frameworks (large frontier developers)

Large frontier developers must “write, implement, comply with, and publish” a frontier AI framework on their website. This framework is the centerpiece of SB-53 compliance. It must describe, at a minimum, how the developer:

• incorporates national and international standards and industry-consensus best practices into its approach to frontier AI safety (for example, NIST’s AI Risk Management Framework and ISO/IEC 42001);
• defines and assesses thresholds for when a model has capabilities that could pose catastrophic risk, potentially using tiered thresholds;
• applies mitigations to address those risks and reviews their adequacy before deployment or extensive internal use;
• uses third-party evaluators to assess catastrophic risk and mitigation effectiveness;
• secures unreleased model weights against unauthorized access or exfiltration;
• identifies and responds to critical safety incidents;
• institutes internal governance practices to ensure the framework is implemented in practice;
• assesses and manages catastrophic risk arising from internal use, including the possibility that a model circumvents internal oversight.

The framework must be reviewed at least annually; any material modification must be republished with a justification within 30 days.

For Daiki, this aligns naturally with our standards-centric approach under ISO/IEC 42001 and the NIST AI RMF. Our existing EU AI Act and quantum governance recipes already require documented, auditable processes for risk identification, mitigation, monitoring, and governance. SB-53 adds a simple rule: if you state a safety process in your framework, you must actually follow it – and you can be sanctioned for misrepresenting that fact.

3.2 Transparency reports on model deployment

All frontier developers – not just large ones – must publish transparency reports whenever they deploy a new frontier model or substantially modify an existing one. These reports must include, at minimum:

• the developer’s website and a human contact mechanism;
• release date of the model;
• supported languages and output modalities;
• intended uses;
• general use restrictions or conditions.

Large frontier developers must additionally summarise:

• catastrophic-risk assessments conducted under their frontier AI framework;
• results of those assessments;
• the role of third-party evaluators;
• other steps taken to satisfy the framework for that specific model.

Developers may integrate these disclosures into system cards or model cards; doing so counts as compliance if the card contains the required information. SB-53 allows narrow redactions to protect trade secrets, cybersecurity, public safety, or U.S. national security, but requires developers to describe the nature and justification of those redactions and retain unredacted versions for five years.

3.3 Critical safety incident reporting and OES

The California Office of Emergency Services (OES) must establish a mechanism for both frontier developers and members of the public to report critical safety incidents.

Frontier developers must:

• report any critical safety incident within 15 days of discovery;
• report incidents posing an imminent risk of death or serious physical injury to appropriate public-safety or law-enforcement authorities within 24 hours.

Beginning 1 January 2027, OES will publish annual anonymised, aggregated reports on critical safety incidents, and the Attorney General will publish aggregated reports on whistleblower disclosures, creating a public evidence base on frontier model incidents.

SB-53 also enables “compliance by reference”: OES may designate federal laws or guidance with equivalent or stricter reporting standards. A frontier developer that opts into such a regime and complies with it is deemed compliant with SB-53’s incident-reporting requirements – but non-compliance with that federal regime also counts as a violation of SB-53.

3.4 Whistleblower protections

SB-53 introduces strong whistleblower safeguards for employees involved in assessing or managing critical safety incidents. A frontier developer may not adopt rules or contracts that prevent covered employees – even anonymously – from disclosing information if they have reasonable cause to believe the developer’s activities pose a specific and substantial danger to public health or safety, or violate the TFAIA (SB-53).

Large frontier developers must provide an internal process through which covered employees can anonymously disclose such information. Retaliation is prohibited, and successful whistleblowers can recover attorney’s fees.

3.5 Enforcement and penalties

Enforcement sits primarily with the California Attorney General:

• Large frontier developers that fail to publish or transmit required documents, misstate catastrophic risk or compliance, fail to report incidents, or fail to follow their own frontier AI framework can face civil penalties of up to USD 1 million per violation.
• Whistleblowers have private rights of action for retaliation.

For developers at or near frontier scale, SB-53 makes AI safety claims legally actionable. A glossy safety page that is not implemented in practice is now an enforcement risk.

4. CalCompute and the standards-first architecture

A distinctive feature of SB-53 is that it links safety obligations to access to compute and to standards.

4.1 CalCompute: a public research cloud

SB-53 establishes CalCompute, a state-backed cloud computing cluster envisioned as public-interest infrastructure for safe, ethical, equitable, and sustainable AI research. A 14-member consortium within the Government Operations Agency will develop a framework for CalCompute and report to the Legislature by 1 January 2027.

CalCompute is meant to:

• foster research and innovation that benefits the public;
• expand access to advanced compute for smaller labs, universities, and civil-society actors;
• embed safety expectations and standards into the conditions of access.

Conceptually, this mirrors the EU AI Act’s emphasis on regulatory sandboxes and testbeds as controlled environments for high-risk AI experimentation with built-in safeguards.

4.2 SB-53 as a standards-centric law

SB-53 repeatedly defers to standards and federal frameworks:

• Frontier AI frameworks must explain how developers incorporate national and international standards and “industry-consensus best practices.”
• The Department of Technology must review, and where appropriate align, definitions of “frontier model”, “frontier developer”, and “large frontier developer” with federal and international thresholds on an annual basis starting in 2027.
• OES may recognise federal reporting regimes as satisfying SB-53’s incident-reporting obligations.

This meta-regulatory design is close to Daiki’s existing narrative: our EU AI Act compliance solution and ISO/IEC 42001 tooling already treat standards, internal management systems, and external regulations as a single, integrated governance stack.

5. SB-53 in a global AI governance landscape

SB-53 arrives into a crowded, fast-evolving global regulatory field:

• The EU AI Act is a horizontal, risk-based framework that applies across sectors and model types, with extensive obligations for high-risk systems and transparency duties for general-purpose models.
• California itself is layering multiple AI measures, including transparency rules for generative systems and pending rules on companion chatbots (e.g. SB-243).
• Other U.S. states, led by Colorado’s AI Act (SB 24-205), are experimenting with high-risk AI regimes that focus on algorithmic discrimination and consequential decisions.

For Daiki clients operating globally, the strategic takeaway is:

SB-53 should not be treated as a standalone compliance silo.

Instead, it is a frontier-AI overlay that can be implemented on top of a robust AI management system (ISO/IEC 42001) and EU AI Act compliance program, using common assets: model registries, AI risk assessments, controls, incident logs, and governance workflows.

6. The Daiki SB-53 Recipe: A practical California SB-53 compliance roadmap

The remainder of this blog focuses on how organisations can operationalise SB-53 – using Daiki tooling and methodology, and activating the dedicated Daiki SB-53 Recipe as the structured implementation layer for frontier models.

Step 1 – Determine if SB-53 applies (now or soon)

• Inventory all foundation models you train or substantially fine-tune, including internal-only models and research systems.
• Estimate training compute (FLOPs) for each. For most organisations, this will require collaboration between model developers, infrastructure teams, and finance.
• Identify any models above or near the 10²⁶ FLOP threshold.
• Assess revenue exposure. If group revenue exceeds USD 500 million, you should plan for the “large frontier developer” obligations.

Daiki’s existing EU AI Act and NIST/ISO 42001 modules already include AI registry and system-classification capabilities that can be adapted to SB-53’s frontier-model concepts.

Step 2 – Build or strengthen your AI governance baseline

Even before you cross the frontier threshold, it is prudent to implement a baseline AI management system aligned with:

• ISO/IEC 42001 (AI management systems),
• the NIST AI Risk Management Framework, and
• the EU AI Act’s technical documentation and post-market monitoring requirements.

This baseline provides clear accountability, end-to-end lifecycle controls, and common templates for risk assessments. At Daiki we use the same approach in our Quantum Governance Recipe – the world’s first QT-QMS. SB-53 can be treated as a domain-specific extension of that governance playbook for frontier AI.

Step 3 – Design a compliant frontier AI framework

For large frontier developers, the frontier AI framework is the central artefact. A practical structure might include:

• Scope and model inventory
• Catastrophic-risk taxonomy
• Capability thresholds and evaluation protocols
• Mitigation strategies and decision criteria
• Third-party evaluation plan
• Cybersecurity of model weights and related infrastructure
• Internal governance (roles, committees, escalation paths)
• Update and review cycle, including triggers for out-of-cycle revisions

The Daiki SB-53 Recipe will map each of these components to concrete tasks, checklists, and evidence artefacts within the platform, ensuring that the framework is not just a PDF but an operational system.

Step 4 – Operationalise transparency reports and model cards

Next, organisations need a repeatable workflow to produce SB-53-compliant transparency reports:

• Standardised model-card templates that capture all mandatory fields (release date, modalities, intended use, restrictions, risk-assessment summaries).
• A redaction protocol that is conservative but defensible, with legal review baked in.
• Integration with your release-management process, so deployment is blocked if required disclosures are incomplete.

Because Daiki already supports EU AI Act technical documentation and ISO 42001 evidence management, SB-53 transparency artefacts can be managed in the same environment, minimising duplicated effort.

Step 5 – Build critical safety incident and whistleblower workflows

SB-53 requires not just reporting channels but end-to-end workflows:

• Detection and triage – how potential critical safety incidents are surfaced from logs, red-team exercises, or user reports.
• Incident assessment – determining whether an event meets the statutory definition (catastrophic risk or critical safety incident).
• Regulatory reporting – filing to OES within 15 days (or 24 hours for imminent threats) and to any other required authorities.
• Post-incident review and CAPA – documenting root causes, corrective and preventive actions, and governance follow-up.

On the whistleblower side, frontier developers should:

• update employee handbooks;
• establish anonymous reporting channels;
• train managers on non-retaliation and escalation;
• log attestations and training completion centrally.

Daiki can orchestrate these workflows, collecting time-stamped evidence suitable for regulators and internal audit.

Step 6 – Harmonise with EU AI Act, ISO 42001, and other regimes

Finally, frontier developers should treat SB-53 as one layer in a global compliance matrix, alongside:

• EU AI Act obligations for high-risk and general-purpose AI;
• ISO/IEC 42001 AI management systems;
• NIST AI RMF risk functions (Map, Measure, Manage, Govern);
• emerging state laws (Colorado SB-205 and others).

Daiki’s platform is designed for exactly this environment: a single AI governance system where evidence, controls, and workflows are reused across multiple regulatory obligations rather than rebuilt for each law.

7. From law to practice: the Daiki SB-53 Recipe

In parallel with this blog, we are rolling out the Daiki SB-53 Recipe – a dedicated use case within the Daiki platform that will:

• implement a frontier model registry with compute-threshold tracking;
• provide frontier AI framework templates aligned with SB-53, ISO/IEC 42001, and the NIST AI RMF;
• generate SB-53-compliant transparency reports and model cards on deployment;
• orchestrate critical safety incident workflows and evidence collection;
• integrate whistleblower policy attestations and training logs;
• map SB-53 artefacts into your broader EU AI Act and quantum/AI governance stack, building on Daiki’s existing EU AI Act solution and Quantum Governance Recipe.

SB-53 is best understood not as an isolated burden but as an opportunity to institutionalise frontier-grade AI safety practices.

As AI systems become deeply embedded in core products, operations, and decision-making, boards and executives will increasingly be judged not on whether they “use AI,” but on whether they can demonstrate disciplined, standards-based AI governance. California SB-53 accelerates this shift by turning responsible AI from a voluntary aspiration into a concrete leadership responsibility that spans risk, strategy, and trust.

Daiki’s mission remains the same: get AI done, responsibly – now with California’s frontier AI transparency law as a concrete, operational benchmark.