California SB 53 — the Transparency in Frontier Artificial Intelligence Act — is the first US state frontier-AI law, in effect since January 1, 2026. Here is exactly who it covers, what it requires, and how to comply.
What is California SB 53?
10^26
FLOP threshold
Training compute (including cumulative fine-tuning) that defines a ‘frontier model.’
$500M
Revenue line
Annual gross revenue above which a frontier developer is a ‘large frontier developer.’
$1M
Max penalty / violation
Civil penalty the California AG can seek, scaled to severity.
California SB 53, formally the Transparency in Frontier Artificial Intelligence Act (TFAIA), is the first US state law written specifically to govern the safety and transparency of the most advanced AI models. Governor Gavin Newsom signed it on September 29, 2025, and its core obligations took effect on January 1, 2026.
The law does not try to regulate every AI system. Instead, California SB 53 targets a narrow band of the largest, most capable models and the well-resourced companies that build them, requiring those developers to publish how they manage catastrophic risk, disclose model capabilities and limitations, report serious safety incidents to the state, and protect employees who raise alarms.
It is the pragmatic successor to SB 1047, the more aggressive bill Newsom vetoed in 2024. Where that bill reached for kill switches and mandatory audits, California SB 53 settles on transparency and reporting — a lighter touch designed to survive the ‘don’t stifle innovation’ objection that killed its predecessor.
California SB 53 (the TFAIA) requires large frontier-AI developers to publish a safety framework and transparency reports, report critical safety incidents to the state within 15 days, and protect whistleblowers — enforced by the Attorney General with penalties up to $1M per violation. In effect since Jan 1, 2026.
Who does California SB 53 apply to?
Two thresholds decide whether California SB 53 touches you, and most companies clear neither.
First, the model must be a frontier model — one trained using more than 10^26 floating-point operations (FLOPs), counting cumulative compute from fine-tuning and later modifications. Second, the heaviest obligations attach only to a large frontier developer: a frontier developer with annual gross revenues above $500 million. That two-part test deliberately aims the law at a handful of the biggest labs while leaving startups and ordinary AI users untouched.
The practical takeaway: if you are deploying or fine-tuning models below that compute bar, or your company is under the revenue line, SB 53’s mandatory duties largely do not apply to you — though the cumulative-compute language means heavy fine-tuning is worth tracking.
| Obligation | All frontier developers | Large frontier developers (>$500M) |
|---|---|---|
| Pre-deployment transparency report | ✓ Required | ✓ Required |
| Annual Frontier AI Framework | — | ✓ Required |
| Critical-incident reporting to Cal OES | ✓ Required | ✓ Required |
| Anonymous whistleblower channel + monthly updates | Basic protections | ✓ Full internal channel required |
What does California SB 53 actually require?
California SB 53 imposes five concrete obligations. Expand each for what it means in practice.
“California SB 53 regulates disclosure, not design — it makes the big labs show their work rather than telling them how to build.”
The core design choice of the TFAIA
1. Transparency report — before deploying a frontier model
Every frontier developer must publish a transparency report before deploying a new or substantially modified frontier model, covering the model’s capabilities, intended uses, limitations, and the results of its risk assessments — including whether third-party evaluators were involved.
2. Frontier AI Framework — annual, large developers only
Large frontier developers must publish an annual framework describing how they identify, mitigate, and govern catastrophic risks: governance structures, cybersecurity measures, and alignment with recognized standards such as the NIST AI Risk Management Framework or ISO/IEC 42001.
3. Critical safety incident reporting — to Cal OES
Developers must notify the California Office of Emergency Services (Cal OES) of a critical safety incident within 15 days of discovery — or within 24 hours if it poses an imminent danger. Covered incidents include unauthorized model tampering, realization of a catastrophic risk, loss of control, or deliberate evasion of safeguards.
4. Whistleblower protections
Developers may not retaliate against employees or contractors who report that the company’s activities pose a specific and substantial danger to public health or safety, or that it violated the TFAIA. Large developers must run an anonymous internal reporting channel and provide monthly status updates to those who use it.
5. CalCompute
SB 53 revives CalCompute, a consortium tasked with building a public cloud-computing cluster to support safe, equitable, and sustainable AI development in the public interest — broadening compute access beyond the largest labs.
How is California SB 53 different from SB 1047?
SB 53 only exists because SB 1047 did not. Newsom vetoed SB 1047 in September 2024 over fears its mandates would stifle innovation, then convened experts whose recommendations shaped the narrower bill that became law.
The difference is philosophical. SB 1047 would have required developers to build ‘kill switches’ into models, conduct prescribed safety testing, and meet audit requirements — design mandates. California SB 53 drops nearly all of that in favor of transparency and reporting: publish your framework, disclose your model, report incidents, protect whistleblowers. It governs what you must reveal, not how you must engineer.
| SB 1047 (vetoed 2024) | California SB 53 (law, 2026) | |
|---|---|---|
| Approach | Design mandates | Transparency & reporting |
| Kill switch / shutdown | Required | Not required |
| Mandatory safety testing & audits | Required | Disclosure of self-run assessments |
| Status | Vetoed Sept 2024 | In effect Jan 1, 2026 |
What are the penalties, and when does California SB 53 take effect?
California SB 53’s obligations have been in force since January 1, 2026. Enforcement runs through the California Attorney General, who can bring civil actions for violations carrying penalties of up to $1 million per violation, scaled to the severity of the offense.
The incident-reporting deadlines are the sharpest operational edge: 15 days from discovery of a critical safety incident, compressed to 24 hours when the incident poses an imminent danger to public safety. For a covered developer, that is a clock that has to live inside the security and on-call process, not the legal department’s quarterly review.
Critical-incident reporting to Cal OES is 15 days from discovery — or 24 hours if there’s imminent danger. Treat it as an operational SLA wired into incident response, not a paperwork afterthought.
How to comply with California SB 53 — a practical checklist
If there is any chance California SB 53 reaches your organization, work this checklist in order. The first step often ends the exercise.
Step 1 — Determine if you’re even in scope
Estimate training compute (including cumulative fine-tuning) against the 10^26 FLOP line, and check annual gross revenue against $500M. Below both? Document the determination and you are largely done. Above the compute line but under the revenue line? You owe transparency reports but not the full framework.
Step 2 — Publish a pre-deployment transparency report
For each new or substantially modified frontier model, document capabilities, intended uses, limitations, and risk-assessment results (and any third-party evaluators) — and publish it before deployment.
Step 3 — (Large developers) Publish your Frontier AI Framework
Map your catastrophic-risk governance to the NIST AI RMF or ISO/IEC 42001, describe cybersecurity and governance structures, and refresh it annually.
Step 4 — Wire up Cal OES incident reporting
Define internally what counts as a critical safety incident, and build the 15-day / 24-hour notification into your incident-response runbook with a named owner.
Step 5 — Stand up whistleblower channels
Publish anti-retaliation notices and, if you are a large developer, an anonymous internal reporting channel with monthly status updates. Keep records.
Builder’s take
Cyntr is nowhere near SB 53’s thresholds — like almost everyone, we build on frontier models rather than train them. But I read California SB 53 as a preview of where the whole stack is heading, and the transparency reports it forces out of the big labs are genuinely useful to downstream builders: they are the closest thing we have to a spec sheet for a model’s real limitations.
- If you fine-tune on top of a frontier base, watch the cumulative-compute language — modifications count toward the 10^26 FLOP line, so ‘I only fine-tuned it’ is not an automatic exemption.
- Even if you are exempt, adopt the transparency-report habit voluntarily: a one-page model card on capabilities, limits, and evals is cheap trust and pre-positions you for the EU AI Act and whatever the US does next.
- The incident-reporting clock (15 days, or 24 hours for imminent danger) is operational, not legal boilerplate — if you are a covered developer, wire it into your on-call runbook now.
- SB 53 is narrow today, but it is the template other states are copying — design your governance to the strictest regime you might plausibly fall under.
Frequently asked questions
Yes. Governor Newsom signed it on September 29, 2025, and its core obligations took effect January 1, 2026. Enforcement is by the California Attorney General.
Almost never. The law only reaches developers of frontier models trained above 10^26 FLOPs, and its heaviest duties apply only to ‘large frontier developers’ with over $500M in annual gross revenue. Startups and ordinary AI users fall outside it.
A model trained using more than 10^26 floating-point operations (FLOPs), counting cumulative compute from fine-tuning and later modifications — a bar only the largest models currently cross.
Incidents such as unauthorized model tampering, realization of a catastrophic risk, loss of control, or deliberate evasion of safeguards. They must be reported to Cal OES within 15 days of discovery, or within 24 hours if there is imminent danger.
The California Attorney General can bring civil actions with penalties of up to $1 million per violation, scaled to the severity of the offense.
They overlap in spirit but differ in scope: SB 53 is a narrow, transparency-focused frontier-model law for the largest US developers, while the EU AI Act is a broad, risk-tiered regime covering many AI uses. A large developer may need to satisfy both.
Primary sources
- Governor Newsom signs SB 53 (Sept 29, 2025) — Governor of California
- SB 53 bill text (official) — California Legislature
- California’s SB 53: The First Frontier AI Law, Explained — Future of Privacy Forum
- What is California’s AI safety law? — Brookings
- Transparency in Frontier Artificial Intelligence Act (SB 53) — standardized AI safety disclosures — WilmerHale
- Transparency in Frontier Artificial Intelligence Act — Wikipedia
Last updated: May 30, 2026. Related: Governance.
