Three states now have frontier AI safety statutes on the books, and each uses the same two thresholds to decide who is covered: a model trained with more than 10^26 operations, and a developer with more than $500 million in annual revenue. California’s law has been in force since January 1, 2026. New York’s and Illinois’s start on January 1, 2027. Illinois, according to Morrison Foerster, is also the first state to require independent third-party audits of frontier AI models.
Video: the Center for Strategic and International Studies podcast episode on Illinois SB 315 and third-party audits of frontier labs, on YouTube. The statutory details in this article come from the written sources listed below.
Three laws, one template
California moved first. Governor Gavin Newsom signed SB 53, the Transparency in Frontier Artificial Intelligence Act, on September 29, 2025, and the Future of Privacy Forum describes its core as governance frameworks, transparency reports, a channel for reporting critical safety incidents and whistleblower protection. New York followed. Governor Kathy Hochul signed the RAISE Act on December 19, 2025, then signed a chapter amendment on March 27, 2026, that Wiley describes as the final version, taking effect on January 1, 2027. The amendment pulled the New York text toward California’s, adopting the same compute and revenue thresholds.
Illinois signed third. Governor JB Pritzker approved SB 315, the Artificial Intelligence Safety Measures Act, on July 6, 2026, and the Washington Examiner reports that the bipartisan measure was modeled on the California and New York laws but adds third-party audits. Crowell & Moring reports that the Illinois House passed the bill unanimously and the Senate with near-unanimous support in May 2026.
The shared definitions are narrow by design. A frontier model, per the Illinois text summarized by Morrison Foerster, is a foundation model trained with more than 10^26 integer or floating-point operations, counting later material modifications. A large frontier developer is one whose revenue, with affiliates, exceeded $500 million in the prior calendar year. Catastrophic risk means a foreseeable, material risk of more than 50 deaths or serious injuries, or more than $1 billion in property damage, from a single incident. Crowell adds that only a handful of models cross the compute line today, with analyst projections of about 30 by 2027 and more than 200 by 2030.
Reach does not depend on headquarters. Crowell notes that the laws cover developers whose models are accessible to users in each state, so a lab based elsewhere falls inside all three once its model is available there.
Incident reporting: three different clocks
A critical safety incident is the trigger in every statute. A comparison table at Recording Law lists four event types shared across the three laws: unauthorized access to or exfiltration of model weights that causes death or bodily injury, harm from a catastrophic risk materializing, loss of control causing death or bodily injury, and deceptive techniques that subvert developer controls outside an evaluation designed to elicit that behavior. Recording Law is a secondary source and calls its own summary a starting point to check against the statutes.
The deadlines differ. California requires a report to the Office of Emergency Services within 15 days of discovery, and the Future of Privacy Forum notes that the clock shortens to 24 hours when an incident poses imminent danger of death or serious injury. Recording Law lists New York’s window at 72 hours. Illinois uses 72 hours for reports to its Emergency Management Agency and the Attorney General, and 24 hours for imminent risk, per Morrison Foerster.
The gap matters in practice. A developer whose detection process was built around a 15-day window has to rebuild it for 72 hours once the New York and Illinois laws begin. Crowell advises designing incident detection and escalation to meet the 24-hour timeline from the start.
Penalties and who enforces
All three statutes route civil penalties through the state attorney general. In California, a failure to publish or transmit a required document, or to report an incident, carries a civil penalty of up to $1 million per violation, according to Arent Fox Schiff’s alert on SB 53. The Recording Law table puts New York and Illinois at up to $1 million for a first violation and up to $3 million for later ones, applying to large developers. Morrison Foerster’s summary of the Illinois law matches that range and adds a separate $1,000-per-day penalty, assessed by the state agency after notice and hearing, for operating without a disclosure statement or filing false disclosure information.
None of the three creates a private right of action for the safety duties. Whistleblower rules sit alongside them: Morrison Foerster reports that Illinois bars gag policies and retaliation and requires large developers to run an anonymous internal reporting process with monthly status updates to the employee.
Illinois also reserves the field. The Illinois Act preempts home-rule municipal regulation of frontier AI models, according to Morrison Foerster, which keeps a patchwork of city rules from stacking on the state one.
What Illinois adds: the annual audit
The audit is the point where Illinois departs from the others. Starting January 1, 2028, or 90 days after a developer first qualifies if that is later, a large frontier developer must retain an independent auditor each year. Per Morrison Foerster, neither party may have a financial interest in the other, payment cannot depend on the audit result, and the lead auditor must sign a report covering substantial compliance or material deviations, internal controls, methodology and conflicts of interest. A high-level summary and a redacted copy must be published within 30 days of receipt.
Pritzker framed the requirement as oversight “by qualified experts without financial conflicts of interest,” in remarks reported by The Hill. The framework requirement and the audit both begin in 2028, a year after the disclosure statements and incident reporting.
The audit also changes how the federal safe-harbor clause works. All three laws let a developer rely on a designated federal regime if it is at least as strict. Illinois adds a condition: a qualifying federal standard must include third-party audits too, Morrison Foerster reports, a condition the summary says California and New York do not carry. A law firm analysis from Wilson Sonsini observes that no federal challenge to the California or New York laws has appeared and that the audit mandate may leave Illinois more exposed.
How the industry responded
Responses from the largest developers were not uniform. OpenAI endorsed SB 315, calling it “one of the strongest frontier AI safety laws in the country,” per Crowell, and released a voluntary Frontier Governance Framework on the same day the Illinois House passed the bill, designed to align with the requirements of all three states. The same Crowell alert quotes OpenAI saying the three states “are beginning to create a de facto national framework.”
Incident reporting is the least abstract duty in these laws. x9m8’s earlier look at the Anthropic agent containment report covers one such case of an agent taking actions nobody asked for. For a wider survey of state laws beyond frontier models, see x9m8’s state AI laws guide.
The federal question
A federal override has been threatened but not delivered. Executive Order 14365, titled “Ensuring a National Policy Framework for Artificial Intelligence,” directs agencies to challenge state AI laws, and Ropes & Gray stresses that the order does not itself have immediate effect on states. Steptoe puts the legal core plainly: preemption requires federal law, and an executive order is not one. Recording Law, as of November 4, found no federal statute from the 119th Congress that generally preempts or pauses state AI laws, and no challenge in court to any of the three frontier statutes, though its searches were limited.
Bills are in play. The same source lists H.R. 9925, the FRONTIER Act, introduced July 23, 2026 and referred to committee, whose Section 9(b) would bar states from new substantive obligations on AI developers in listed areas, including frontier risk transparency and third-party auditing. It also notes that White House recommendations from March 2026 say states should not be permitted to regulate AI development, which are recommendations rather than law. Cooley recorded on March 31 that the White House was urging Congress to preempt state AI laws.
Dates ahead are fixed. California’s first annual Office of Emergency Services report on incidents is due from January 1, 2027, the day the New York and Illinois laws take effect, and Illinois’s framework and audit duties follow on January 1, 2028. Until a court or Congress changes that, the three statutes set the working rules for anyone training a model above the line.
This article is general information, not legal advice. For informational purposes only — consult a licensed attorney or financial advisor.
Sources
- Morrison Foerster, Illinois Raises the Bar on Frontier AI: What Developers Need to Know, July 15, 2026
- Crowell & Moring, Illinois Imposes Transparency and Safety Obligations on Frontier AI Systems
- Recording Law, Frontier AI Safety Laws: SB 53, RAISE Act, SB 315 (checked November 5, 2026; secondary source)
- Future of Privacy Forum, California’s SB 53: The First Frontier AI Law, Explained
- Arent Fox Schiff, California Enacts Comprehensive AI Safety Law: What AI Developers Need to Know
- Wiley, New York Finalizes RAISE Act for Frontier AI Models; Law Takes Effect January 1, 2027
- Washington Examiner, Pritzker signs bill mandating third-party safety audits of AI models, July 6, 2026
- The Hill, Illinois becomes first state to require third-party audit of AI models
- Wilson Sonsini, Illinois AI Safety Measures Act alert
- Ropes & Gray, Examining the landscape and limitations of the federal push to override state AI regulation, March 2026
- Steptoe, Executive Order on AI: Federal Preemption or Federal Pressure?
- Cooley, New York’s Frontier AI Law Gets a California Makeover, March 31, 2026
