CHAPTER 22.6. Safe and Secure Innovation for Frontier Artificial Intelligence Systems

22602.

As used in this chapter:

(a) “Advanced persistent threat” means an adversary with sophisticated levels of expertise and significant resources that allow it, through the use of multiple different attack vectors, including, but not limited to, cyber, physical, and deception, to generate opportunities to achieve its objectives that are typically to establish and extend its presence within the information technology infrastructure of organizations for purposes of exfiltrating information or to undermine or impede critical aspects of a mission, program, or organization or place itself in a position to do so in the future.

(b) “Artificial intelligence model” means a machine-based system that can make predictions, recommendations, or decisions influencing real or virtual environments and can use model inference to formulate options for information or action.

(c) “Artificial intelligence safety incident” means any of the following:

(1) A covered model autonomously engaging in a sustained sequence of unsafe behavior other than at the request of a user.

(2) Theft, misappropriation, malicious use, inadvertent release, unauthorized access, or escape of the model weights of a covered model.

(3) The critical failure of technical or administrative controls, including controls limiting the ability to modify a covered model, designed to limit access to a hazardous capability of a covered model.

(4) Unauthorized use of the hazardous capability of a covered model.

(d) “Computing cluster” means a set of machines transitively connected by data center networking of over 100 gigabits that has a theoretical maximum computing capacity of 10^20 integer or floating-point operations per second for training artificial intelligence.

(e) “Covered guidance” means any of the following:

(1) Applicable guidance issued by the National Institute of Standards and Technology and by the Frontier Model Division.

(2) Industry best practices, including relevant safety practices, precautions, or testing procedures undertaken by developers of comparable models, and any safety standards or best practices commonly or generally recognized by relevant experts in academia or the nonprofit sector.

(3) Applicable safety-enhancing standards set by standards setting organizations.

(f) “Covered model” means an artificial intelligence model that meets either of the following criteria:

(1) The artificial intelligence model was trained using a quantity of computing power greater than 10^26 integer or floating-point operations in 2024, or a model that could reasonably be expected to have similar performance on benchmarks commonly used to quantify the performance of state-of-the-art foundation models, as determined by industry best practices and relevant standard setting organizations.

(2) The artificial intelligence model has capability below the relevant threshold on a specific benchmark but is of otherwise similar general capability.

(g) “Critical harm” means a harm listed in paragraph (1) of subdivision (n).

(h) “Critical infrastructure” means assets, systems, and networks, whether physical or virtual, the incapacitation or destruction of which would have a debilitating effect on physical security, economic security, public health, or safety in the state.

(i) (1) “Derivative model” means an artificial intelligence model that is a derivative of another artificial intelligence model, including either of the following:

(A) A modified or unmodified copy of an artificial intelligence model.

(B) A combination of an artificial intelligence model with other software.

(2) “Derivative model” does not include an entirely independently trained artificial intelligence model.

(j) (1) “Developer” means a person that creates, owns, or otherwise has responsibility for an artificial intelligence model.

(2) “Developer” does not include a third-party machine-learning operations platform, an artificial intelligence infrastructure platform, a computing cluster, an application developer using sourced models, or an end-user of an artificial intelligence model.

(k) “Fine tuning” means the adjustment of the model weights of an artificial intelligence model that has been previously trained by training the model with new data.

(l) “Frontier Model Division” means the Frontier Model Division created pursuant to Section 11547.6 of the Government Code.

(m) “Full shutdown” means the cessation of operation of a covered model, including all copies and derivative models, on all computers and storage devices within custody, control, or possession of a person, including any computer or storage device remotely provided by agreement.

(n) (1) “Hazardous capability” means the capability of a covered model to be used to enable any of the following harms in a way that would be significantly more difficult to cause without access to a covered model:

(A) The creation or use of a chemical, biological, radiological, or nuclear weapon in a manner that results in mass casualties.

(B) At least five hundred million dollars ($500,000,000) of damage through cyberattacks on critical infrastructure via a single incident or multiple related incidents.

(C) At least five hundred million dollars ($500,000,000) of damage by an artificial intelligence model that autonomously engages in conduct that would violate the Penal Code if undertaken by a human.

(D) Other threats to public safety and security that are of comparable severity to the harms described in paragraphs (A) to (C), inclusive.

(2) “Hazardous capability” includes a capability described in paragraph (1) even if the hazardous capability would not manifest but for fine tuning and posttraining modifications performed by third-party experts intending to demonstrate those abilities.

(o) “Machine-learning operations platform” means a solution that includes a combined offering of necessary machine-learning development capabilities, including exploratory data analysis, data preparation, model training and tuning, model review and governance, model inference and serving, model deployment and monitoring, and automated model retraining.

(p) “Model weight” means a numerical parameter established through training in an artificial intelligence model that helps determine how input information impacts a model’s output.

(q) “Open-source artificial intelligence model” means an artificial intelligence model that is made freely available and may be freely modified and redistributed.

(r) “Person” means an individual, proprietorship, firm, partnership, joint venture, syndicate, business trust, company, corporation, limited liability company, association, committee, or any other nongovernmental organization or group of persons acting in concert.

(s) “Positive safety determination” means a determination, pursuant to subdivision (a) or (c) of Section 22603, with respect to a covered model that is not a derivative model that a developer can reasonably exclude the possibility that a covered model has a hazardous capability or may come close to possessing a hazardous capability when accounting for a reasonable margin for safety and the possibility of posttraining modifications.

(t) “Posttraining modification” means the modification of the capabilities of an artificial intelligence model after the completion of training by any means, including, but not limited to, initiating additional training, providing the model with access to tools or data, removing safeguards against hazardous misuse or misbehavior of the model, or combining the model with, or integrating it into, other software.

(u) “Safety and security protocol” means documented technical and organizational protocols that meet both of the following criteria:

(1) The protocols are used to manage the risks of developing and operating covered models across their life cycle, including risks posed by enabling or potentially enabling the creation of derivative models.

(2) The protocols specify that compliance with the protocols is required in order to train, operate, possess, and provide external access to the developer’s covered model.

22603.

(a) Before initiating training of a covered model that is not a derivative model, a developer of that covered model shall determine whether it can make a positive safety determination with respect to the covered model.

(1) In making the determination required by this subdivision, a developer shall incorporate all covered guidance.

(2) A developer may make a positive safety determination if the covered model will have lower performance on all benchmarks relevant under subdivision (f) of Section 22602 than either of the following:

(A) A non-covered model that manifestly lacks hazardous capabilities.

(B) Another model that is the subject of a positive safety determination.

(3) Upon making a positive safety determination, the developer of the covered model shall submit to the Frontier Model Division a certification under penalty of perjury that specifies the basis for that conclusion.

(b) Before initiating training of a covered model that is not a derivative model that is not the subject of a positive safety determination, and until that covered model is the subject of a positive safety determination, the developer of that covered model shall do all of the following:

(1) Implement administrative, technical, and physical cybersecurity protections to prevent unauthorized access to, or misuse or unsafe modification of, the covered model, including to prevent theft, misappropriation, malicious use, or inadvertent release or escape of the model weights from the developer’s custody, that are appropriate in light of the risks associated with the covered model, including from advanced persistent threats or other sophisticated actors.

(2) Implement the capability to promptly enact a full shutdown of the covered model.

(3) Implement all covered guidance.

(4) Implement a written and separate safety and security protocol that does all of the following:

(A) Provides reasonable assurance that if a developer complies with its safety and security protocol, either of the following will apply:

(i) The developer will not produce a covered model with a hazardous capability or enable the production of a derivative model with a hazardous capability.

(ii) The safeguards enumerated in the policy will be sufficient to prevent critical harms from the exercise of a hazardous capability in a covered model.

(B) States compliance requirements in an objective manner and with sufficient detail and specificity to allow the developer or a third party to readily ascertain whether the requirements of the safety and security protocol have been followed.

(C) Identifies specific tests and test results that would be sufficient to reasonably exclude the possibility that a covered model has a hazardous capability or may come close to possessing a hazardous capability when accounting for a reasonable margin for safety and the possibility of posttraining modifications, and in addition does all of the following:

(i) Describes in detail how the testing procedure incorporates fine tuning and posttraining modifications performed by third-party experts intending to demonstrate those abilities.

(ii) Describes in detail how the testing procedure incorporates the possibility of posttraining modifications.

(iii) Describes in detail how the testing procedure incorporates the requirement for reasonable margin for safety.

(iv) Provides sufficient detail for third parties to replicate the testing procedure.

(D) Describes in detail how the developer will meet requirements listed under paragraphs (1), (2), (3), and (5).

(E) If applicable, describes in detail how the developer intends to implement the safeguards and requirements referenced in paragraph (1) of subdivision (d).

(F) Describes in detail the conditions that would require the execution of a full shutdown.

(G) Describes in detail the procedure by which the safety and security protocol may be modified.

(H) Meets other criteria stated by the Frontier Model Division in guidance to achieve the purpose of maintaining the safety of a covered model with a hazardous capability.

(5) Ensure that the safety and security protocol is implemented as written, including, at a minimum, by designating senior personnel responsible for ensuring implementation by employees and contractors working on a covered model, monitoring and reporting on implementation, and conducting audits, including through third parties as appropriate.

(6) Provide a copy of the safety and security protocol to the Frontier Model Division.

(7) Conduct an annual review of the safety and security protocol to account for any changes to the capabilities of the covered model and industry best practices and, if necessary, make modifications to the policy.

(8) If the safety and security protocol is modified, provide an updated copy to the Frontier Model Division within 10 business days.

(9) Refrain from initiating training of a covered model if there remains an unreasonable risk that an individual, or the covered model itself, may be able to use the hazardous capabilities of the covered model, or a derivative model based on it, to cause a critical harm.

(c) (1) Upon completion of the training of a covered model that is not the subject of a positive safety determination and is not a derivative model, the developer shall perform capability testing sufficient to determine whether the developer can make a positive safety determination with respect to the covered model pursuant to its safety and security protocol.

(2) Upon making a positive safety determination with respect to the covered model, a developer of the covered model shall submit to the Frontier Model Division a certification of compliance with the requirements of this section within 90 days and no more than 30 days after initiating the commercial, public, or widespread use of the covered model that includes both of the following:

(A) The basis for the developer’s positive safety determination.

(B) The specific methodology and results of the capability testing undertaken pursuant to this subdivision.

(d) Before initiating the commercial, public, or widespread use of a covered model that is not subject to a positive safety determination, a developer of the nonderivative version of the covered model shall do all of the following:

(1) Implement reasonable safeguards and requirements to do all of the following:

(A) Prevent an individual from being able to use the hazardous capabilities of the model, or a derivative model, to cause a critical harm.

(B) Prevent an individual from being able to use the model to create a derivative model that was used to cause a critical harm.

(C) Ensure, to the extent reasonably possible, that the covered model’s actions and any resulting critical harms can be accurately and reliably attributed to it and any user responsible for those actions.

(2) Provide reasonable requirements to developers of derivative models to prevent an individual from being able to use a derivative model to cause a critical harm.

(3) Refrain from initiating the commercial, public, or widespread use of a covered model if there remains an unreasonable risk that an individual may be able to use the hazardous capabilities of the model, or a derivative model based on it, to cause a critical harm.

(e) A developer of a covered model shall periodically reevaluate the procedures, policies, protections, capabilities, and safeguards implemented pursuant to this section in light of the growing capabilities of covered models and as is reasonably necessary to ensure that the covered model or its users cannot remove or bypass those procedures, policies, protections, capabilities, and safeguards.

(f) (1) A developer of a nonderivative covered model that is not the subject of a positive safety determination shall submit to the Frontier Model Division an annual certification of compliance with the requirements of this section signed by the chief technology officer, or a more senior corporate officer, in a format and on a date as prescribed by the Frontier Model Division.

(2) In a certification submitted pursuant to paragraph (1), a developer shall specify or provide, at a minimum, all of the following:

(A) The nature and magnitude of hazardous capabilities that the covered model possesses or may reasonably possess and the outcome of capability testing required by subdivision (c).

(B) An assessment of the risk that compliance with the safety and security protocol may be insufficient to prevent harms from the exercise of the covered model’s hazardous capabilities.

(C) Other information useful to accomplishing the purposes of this subdivision, as determined by the Frontier Model Division.

(g) A developer shall report each artificial intelligence safety incident affecting a covered model to the Frontier Model Division in a manner prescribed by the Frontier Model Division. The notification shall be made in the most expedient time possible and without unreasonable delay and in no event later than 72 hours after learning that an artificial intelligence safety incident has occurred or learning facts sufficient to establish a reasonable belief that an artificial intelligence safety incident has occurred.

(h) (1) Reliance on an unreasonable positive safety determination does not relieve a developer of its obligations under this section.

(2) A positive safety determination is unreasonable if the developer does not take into account reasonably foreseeable risks of harm or weaknesses in capability testing that lead to an inaccurate determination.

(3) A risk of harm or weakness in capability testing is reasonably foreseeable, if, by the time that a developer releases a model, an applicable risk of harm or weakness in capability testing has already been identified by either of the following:

(A) Any other developer of a comparable or comparably powerful model through risk assessment, capability testing, or other means.

(B) By the United States Artificial Intelligence Safety Institute, the Frontier Model Division, or any independent standard-setting organization or capability-testing organization cited by either of those entities.

22604.

A person that operates a computing cluster shall implement appropriate written policies and procedures to do all of the following when a customer utilizes compute resources that would be sufficient to train a covered model:

(a) Obtain a prospective customer’s basic identifying information and business purpose for utilizing the computing cluster, including all of the following:

(1) The identity of that prospective customer.

(2) The means and source of payment, including any associated financial institution, credit card number, account number, customer identifier, transaction identifiers, or virtual currency wallet or wallet address identifier.

(3) The email address and telephonic contact information used to verify a prospective customer’s identity.

(4) The Internet Protocol addresses used for access or administration and the date and time of each access or administrative action.

(b) Assess whether a prospective customer intends to utilize the computing cluster to deploy a covered model.

(c) Annually validate the information collected pursuant to subdivision (a) and conduct the assessment required pursuant to subdivision (b).

(d) Maintain for seven years and provide to the Frontier Model Division or the Attorney General, upon request, appropriate records of actions taken under this section, including policies and procedures put into effect.

(e) Implement the capability to promptly enact a full shutdown in the event of an emergency.

22605.

(a) A developer of a covered model that provides commercial access to that covered model shall provide a transparent, uniform, publicly available price schedule for the purchase of access to that covered model at a given level of quality and quantity subject to the developer’s terms of service and shall not engage in unlawful discrimination or noncompetitive activity in determining price or access.

(b) A person that operates a computing cluster shall provide a transparent, uniform, publicly available price schedule for the purchase of access to the computing cluster at a given level of quality and quantity subject to the developer’s terms of service and shall not engage in unlawful discrimination or noncompetitive activity in determining price or access.

22606.

(a) If the Attorney General has reasonable cause to believe that a person is violating this chapter, the Attorney General shall commence a civil action in a court of competent jurisdiction.

(b) In a civil action under this section, the court may award any of the following:

(1) (A) Preventive relief, including a permanent or temporary injunction, restraining order, or other order against the person responsible for a violation of this chapter, including deletion of the covered model and the weights utilized in that model.

(B) Relief pursuant to this paragraph shall be granted only in response to harm or an imminent risk or threat to public safety.

(2) Other relief as the court deems appropriate, including monetary damages to persons aggrieved and an order for the full shutdown of a covered model.

(3) A civil penalty in an amount not exceeding 10 percent of the cost, excluding labor cost, to develop the covered model for a first violation and in an amount not exceeding 30 percent of the cost, excluding labor cost, to develop the covered model for any subsequent violation.

(c) In the apportionment of penalties assessed pursuant to this section, defendants shall be jointly and severally liable.

(d) A court shall disregard corporate formalities and impose joint and several liability on affiliated entities for purposes of effectuating the intent of this section if the court concludes that both of the following are true:

(1) Steps were taken in the development of the corporate structure among affiliated entities to purposely and unreasonably limit or avoid liability.

(2) The corporate structure of the developer or affiliated entities would frustrate recovery of penalties or injunctive relief under this section.

22607.

(a) Pursuant to subdivision (a) of Section 1102.5 of the Labor Code, a developer shall not prevent an employee from disclosing information to the Attorney General if the employee has reasonable cause to believe that the information indicates that the developer is out of compliance with the requirements of Section 22603.

(b) Pursuant to subdivision (b) of Section 1102.5 of the Labor Code, a developer shall not retaliate against an employee for disclosing information to the Attorney General if the employee has reasonable cause to believe that the information indicates that the developer is out of compliance with the requirements of Section 22603.

(c) The Attorney General may publicly release any complaint, or a summary of that complaint, pursuant to this section if the Attorney General concludes that doing so will serve the public interest.

(d) Employees shall seek relief for violations of this section pursuant to Sections 1102.61 and 1102.62 of the Labor Code.

(e) Pursuant to subdivision (a) of Section 1102.8 of the Labor Code, a developer shall provide clear notice to all employees working on covered models of their rights and responsibilities under this section.

safety articles